COMPANY TO DEAL WITH A MAJOR SERVICE CRISIS
An exploratory study
Master Thesis Joost Marico 1
University of Groningen, the Netherlands 26 - 06 - 2013
http://hananial.com/pix/recovery.jpg
The best response strategy for a service company to
deal with a major service crisis
An exploratory study
Master Thesis
Master of Business Administration: Marketing Management Faculty of Economics and Business, University of Groningen Date of completion 26-06-2013
First supervisor: dr. ir. Maarten Gijsenberg Second supervisor: N. (Niels) Holtrop, MSc.
Joost Marico
Tweede Willemstraat 41A 9725 JH Groningen The Netherlands
i
ABSTRACT
This explorative study examines which action and communication response strategies are best to implement by service companies that face a major crisis/failure. Firms that are facing such crises often react in a rather clumsy way to deal with the problem. Research has been done on recovery strategies, however, most relevant studies do not focus on the importance of the different service recovery attributes. These attributes have a moderating effect on the negative relation that a service failure has on customer satisfaction. The contribution of this study is that a general theoretical framework is built to research which of these recovery attributes with its levels are found to be most important by customers. Choice based conjoint analysis with latent class analysis is used for measuring the preferences of respondents.
ii
PREFACE
This thesis forms the final part of my Master of Science degree in Business Administration with the specialization Marketing Management. Next to management courses, I followed useful marketing intelligence courses. This enabled me to complete this thesis, which has a more research kind of topic. I enjoyed writing this thesis more than I expected beforehand. This thesis is a nice ending of a great time being a student at the University of Groningen.
First of all, I would like to thank my supervisor Maarten Gijsenberg for his time and his detailed and constructive feedback. Furthermore, I would like to thank my fellow students for providing help when needed. Moreover, I really want to thank my friends, family, fellow students, acquaintances and other unknown respondents for filling in my survey. Without them, finishing this thesis was not possible. Lastly, I want to thank my parents, brother and roommates who supported me during this thesis.
Joost Marico
iii
TABLE OF CONTENTS
1. INTRODUCTION ... 1
1.1 Theoretical, managerial and social relevance ... 2
1.2 Structure of thesis ... 3
2. THEORETICAL FRAMEWORK ... 4
2.1 Definitions and relations ... 4
2.1.1 Service failure ... 4
2.1.2 Types of service failure ... 4
2.1.3 Customer satisfaction and loyalty ... 5
2.1.4 Service failure and customer satisfaction ... 5
2.1.5 Service recovery (and its evolution) ... 6
2.1.6 Service recovery and customer satisfaction ... 7
2.1.7 Service recovery paradox ... 8
2.2 Introducing the variables ... 8
2.2.1 Moderating role of recovery strategies ... 8
2.2.2 Different media channels ... 10
2.2.3 Five service recovery variables ... 11
2.2.4 Socio-demographic variables ... 12
2.3 Conceptual Model ... 13
3. RESEARCH DESIGN ... 15
3.1 Research method: CBC analysis... 15
3.2 Study design of CBC ... 15
3.2.1 Number of attributes and levels ... 15
3.2.2 Design of choice task ... 17
3.2.3 Holdout task ... 18
3.3 Experimental procedure ... 18
3.4. Plan of analysis ... 19
4. ANALYSIS AND RESULTS ... 21
4.1 Sample description ... 21
4.2 Latent class analysis – Aggregate level ... 22
4.3 Latent Class analysis - Segment level ... 25
4.4 Reliability check ... 29
5. CONCLUSION ... 30
5.1 Discussion results ... 30
5.2 Managerial recommendations ... 32
5.2 Limitations and recommendations for further research ... 34
6. REFERENCES ... 36
1
1. INTRODUCTION
Chances are high for consumers that they have experienced multiple product or service failures by firms in their lifetime. These failures could be small, but also large which can lead to inconveniences for customers. In case of a product-harm crisis, which can be defined as events wherein products are found to be defective or even dangerous (Dawar and Pillutla 2000), one can think of products that do not meet their standards. This means that they fail in delivering the promised quality and that they are easily broken, or that they contain parts that could form a threat to our health. In case of a service crisis, one can think of long waiting times or no mobile phone connection. To summarize, when firms are experiencing such product or service crises they are not able to meet consumers’ expectations.
Unfortunately, firms in the marketplace have a chance to be confronted with these product or service crises, even a large firm like Vodafone. In April 2012, Vodafone faced a major critical service failure in the Netherlands that lasted for several days (Noordhuis 2012). An important building caught fire with the consequence that many customers could not use their Vodafone network. For some customers this was definitely not a pleasant experience, especially for people who are dependent on their mobile phone. More recently in September 2012 and in the beginning of 2013, Vodafone had another major service failure; only many blackberry phone owners were hit (Tritel.nl). In addition, Blackberry faced a major service failure in 2011 where it took three days to solve the problems (Kraan 2012). As compensation, consumers were given free apps worth around 100 dollars. Furthermore, the Dutch banking firm ING faced multiple major service crises with their online banking system. For example, this was the case in the beginning of 2012 (van Hoek 2012). Customers were not able to login at the bank for some days in a row, which was pretty inconvenient for most customers. Especially for business customers who rely on the online banking system for their webshops for example. To give another example, the Dutch railroad company (the Nederlandse Spoorwegen (NS)) also had to deal with a lot of major service crises due to bad weather conditions during the winter. This resulted in delayed trains or no trains at all. As a consequence, many people were late for work or school, or had to depart earlier to arrive in time.
2
dissatisfied customers told others about their bad experience. Furthermore, when the number of consumers experiencing dissatisfaction is high enough, these responses may have lasting effects in terms of negative image and reduced sales for the company. Today in the online world, firms have to be vigilant since word-of-mouth can be spread more easily.
Firms that are facing such product and service crises often react in a rather clumsy way to deal with the problems which is reflected by their communication or action response strategy. This can lead to customers getting more angry and frustrated, and can decrease customer satisfaction. Research by Berry and Parasuraman (1991) has indicated that only 50-67% of customers who experienced difficulties with one of five service companies were satisfied with the outcome. This indicates that firms have a lot to gain in this area. However, a service failure can also have a positive side, since research by Smith and Bolton (1998) has shown that a good handling of complaints in individual service encounters may actually pay off in the end. They also find support for the existence of a service recovery paradox, which means that a highly satisfactory recovery will maintain or even increase cumulative satisfaction and loyalty. The problem addressed in this study is the clumsy way by which firms sometimes react to major service failures. The purpose of this research is to investigate how companies can best deal with major general service crises. Therefore, the research questions of the study at hand are the following:
‘What is the best action response strategy for a company to result in the best outcome of a major service crisis?’
‘What is the best communication response strategy for a company to result in the best outcome of a major service crisis?’
1.1 Theoretical, managerial and social relevance
3
customers. The service recovery attributes that are valued most will also have a more positive moderate effect on the negative relation that service failure has on customer satisfaction. In addition, this study only focuses on major service crises. The reason for this is that the service industry is becoming more and more important (Bitner, Booms and Tetreault 1990; Spreng, Harrell and Mackoy 1995). Moreover, research on major general service crises is limited since most literature is on the individual service encounter level, despite the major impact. Another reason is to make the research more concrete and researchable.
In this study, a large mobile telecom provider is taken as an example. When a mobile telecom provider experiences such a major service crisis, many customers will suffer from it since most customers are very dependent on their mobile phone connection. The results of this study are beneficial for customers when they are properly implemented by service companies, since it will minimize the decrease in customer satisfaction after a service crisis (or even increase satisfaction). Satisfied customers are in turn beneficial for the company since customer satisfaction is linked to profits (Sparks and McColl-Kennedy 2001). Lastly, this study distinguishes between action and communication strategies which makes it practical for managers to implement.
1.2 Structure of thesis
4
2. THEORETICAL FRAMEWORK
2.1 Definitions and relations
2.1.1 Service failure
The concept of service failure is well researched in literature. Service failures are common in the service industry (McColl-Kennedy, Daus and Sparks 2003) and according to Hart, Hesket and Sasser (2000) the fact is that they appear to be inevitable for companies. A simple and clear definition for a service failure is as follows: a service performance that falls below a customer’s expectations (Bell and Zemke 1987; Hoffman and Bateson 1997). Spreng, Harrel and Mackoy (1995) describe a service failure as a flawed outcome that reflects a breakdown in reliability. A service failure can occur anytime during the customer’s relationship with a service organization (Kelley 1994). At an early stage it will damage the customer’s overall evaluation of the organization more since fewer successful service experiences with that organization are lacking that can counterbalance the failure (Boulding et al. 1993). Service failures can differ in respect to timing, severity, and frequency dimensions (Kelley 1994). In the introduction is made clear that this thesis is focusing on a service crisis and not on a product-harm crisis. In the literature this distinction can also be made, since it has been suggested that a service failure is profoundly different from the failure of a tangible product (Hocutt 1997). An explanation for this view is that a service usually has a psychological and largely personal outcome, whereas a tangible product failure usually is perceived as more impersonal in its impact on a customer.
2.1.2 Types of service failure
5
2.1.3 Customer satisfaction and loyalty
A considerable amount of evidence is gathered to support the notion that customer satisfaction is vital to the success of organizations. A reason for this is that customer satisfaction is linked to profits (Sparks and McColl-Kennedy 2001) and is important in both the short and long term (Miller, Craighead and Karwan 2000). According to Westbrook (1980), customer satisfaction refers to an individual’s subjectively derived favorable evaluation of any outcome and/or experience associated with consuming a product. According to Maxham (2001), researchers have multiple perspectives on satisfaction. It can be seen as a purchase outcome whereby consumers compare rewards and costs with anticipated consequences. Another perspective is more operational, where satisfaction is more similar to attitude, as it represents the sum of several attribute satisfaction judgments. This perspective is a transaction specific measure (Bitner 1990; Parasuraman, Zeithaml and Berry 1988). Other researchers like Cronin and Taylor (1994) view satisfaction as a cumulative evaluation, and an outgrowth of service quality. Here, satisfaction is more a global judgment rather than a transaction-specific measure. In addition, satisfaction is seen to have an affective element that is experiential and that can be judged best after consumption (Ostrom and Lacobucci 1995). Furthermore, the satisfaction literature suggests that increased satisfaction with a service encounter leads likely to returning to the same service provider (Harris 2006). Several researchers (e.g., Bitner 1990; Boulding et al. 1993) state that service encounter satisfaction has greater perceived service quality as a consequence, which results in service loyalty. Besides the customer satisfaction outcome measure, measures like customer loyalty and retention are also typical measures to use to see if an organization is successful (Miller, Craighead and Karwan 2000). Compared to customer satisfaction, these measures are the long-term goals of most service managers. In addition, van Doorn and Verhoef (2008) indicate that early studies researched the process of satisfaction formation, while recent research is more focusing on the relationship between customer satisfaction and customer loyalty.
2.1.4 Service failure and customer satisfaction
6
customer satisfaction is. Moreover, Keaveney (1995) states that service failure is the core reason for service switching behavior of consumers. This indicates the importance of well-executed service recoveries for enhancing customer satisfaction, building customer relationships, and preventing customer defections (Smith, Bolton and Wagner 1999).
2.1.5 Service recovery (and its evolution)
The concept of service recovery has evolved over time. In the early 1970s and 1980s the term was more used for restoring and recovering from natural disasters (Brown, Cowles and Tuten 1996). However, from late 1970s marketers began to see the importance of the long-term benefits of recovery such as improved customer loyalty and favorable word-of-mouth communication, besides only resolving service problems (Brown, Cowles and Tuten 1996). In 1990, Hart, Hesker and Sasser published an article which caused attention to shift to the proactive, strategic role that service recovery can play in a competitive marketplace. Many definitions of service recovery can be found in the literature. The most common definitions will be presented here. Service recovery has been defined as:
Satisfactory problem resolution (Brown, Cowles and Tuten 1996)
The actions a service provider takes to respond to a service failure (Brown, Cowles and Tuten 1996)
Doing the service right the second time (Brown, Cowles and Tuten 1996)
7
Table 1: Overview of studies about service recovery
Service recovery topic Research
Service recovery’s multi-dimensional nature Johnston (1994); Davidow (2003) Antecedents service recovery Kelley and Davis (1994)
Outcomes of service recovery McColl-Kennedy and Sparks (2003) Outcomes of service recovery (theoretical
perspective)
Andreassen (1999); Levesque and McDougall (2000); Bougi, Pieters and Zeelenberg. (2003)
Outcomes of service recovery (empirical perspective)
Narayanda (1998); McCollough, Berry and Yadav (2000); Smith and Bolton (1998)
Complaining behavior Stephens and Gwinner (1998); Andreasen
(1988) or Davidow (2003) Influence of competitive environment on service
recovery efforts
Estelami (2000) Relationship between satisfaction with service
recovery and cumulative satisfaction
McCollough, Berry and Yadav (2000) Role of fairness theory, equity theory and justice Oliver and Swan (1989); de Ruyter and
Wetzels (2000); McColl-Kennedy and Sparks (2003)
Relationship between service recovery perceptions of justice
Saxby, Tat and Johansen (2000); Blodgett, Hill and Tax (2001)
Impact of relationship factors on satisfaction with service performance after recovery”
Hess, Ganesan and Klein (2003)
2.1.6 Service recovery and customer satisfaction
8
empowerment of the employee, feedback about the progress to the customer, atonement (apology), explanation and tangibles (appearance of employees and environment). These dimensions by Boshoff (2005) were established by research in the financial sector, therefore, he states that the results cannot be generalized to all service providers. In addition, the development of this instrument was primarily based on the disconfirmation paradigm (Boshoff 2005), which is discussed next.
2.1.7 Service recovery paradox
A service recovery related concept is the ‘service recovery paradox’, which means that customer attitudes will be better after a service recovery than they were prior to any service failure, given a proper service recovery effort (Maxham 2001). This paradox can be explained by the expectancy disconfirmation theory and attribution theory.
To start, according to the disconfirmation literature, a positive disconfirmation is the result when an actual outcome exceeds a customer’s expected outcome. Vice versa when expectation are not met (Bearden and Teel 1983; Oliver 1980). The paradox can be explained since consumers first compare expectations for recovery to their perceptions of service recovery performance (Maxham 2001). In case of the paradox, recovery expectations are positively disconfirmed, which leads to higher perceptions of satisfaction. It is important for service companies to know what customers actually expect in response to service failures. Once known, it is possible to measure service recovery performance, and indirectly also customer satisfaction relative to the expectations (Boshoff 1999). Secondly, the so-called attribution theory may provide further insight into the paradox, since it suggests that consumers think after the first service failure that it was a onetime incident or that the failure is beyond the company’s control (i.e., an unstable attribution). This leads to more positive feelings towards that company and can trigger a recovery paradox. However, if another service failure follows, complainants may discount the circumstantial attribution and instead believe that the company consistently fails (i.e., a stable attribution) and they reevaluate their attributions (Maxham 2001; 2002). Customers will likely infer that multiple failures are due to problems inherent to the firm (Maxham 2002). When this is the case, the customer will only be more dissatisfied.
2.2 Introducing the variables
2.2.1 Moderating role of recovery strategies
9
is relevant, which begins when the company becomes aware of a failure and ends when a fair restitution has been made to the customer (Miller, Craighead and Karwan 2000). When this phase takes too long, research demonstrates that a significant decline in loyalty and satisfaction will be the result (Spreng, Harrell and Mackoy 1995). Furthermore, Miller, Craighead and Karwan (2000) distinguish different types of service recovery activities which can be divided in a psychological and a tangible form.
Psychological recovery efforts are focused on improving the situation by showing concern. Most used techniques are empathy and apology (Miller, Craighead and Karwan 2000). These techniques are simple and inexpensive, and are powerful when used together. However, caution is advised since inappropriately usage like a non-empathetic apology can be worse than none at all (Miller, Craighead and Karwan 2000). Another study indicates that apology alone was the least effective strategy, however when used with compensation and assistance customer loyalty improved significantly (Levesque and McDougall 2000). Tangible recovery efforts comprise forms of compensation for the costs and inconveniences caused by the failure, which can be higher than the fair remedy to atone for a bad experience (Miller, Craighead and Karwan 2000). However, according to Smith, Bolton and Wagner (1999) caution is advised since higher compensation may lead consumers to feel less satisfied because of ‘over-rewarding’. This means an upper limit to over-compensation is advised. Furthermore, Miller, Craighead and Karwan (2000) mention two types of delivery of service recovery: front-line empowerment and speed of recovery. The first states that if employees have the right knowledge and power to compensate a dissatisfied customer, chance is higher that a dissatisfied customer will become satisfied and retains (Miller, Craighead and Karwan 2000). The speed of recovery will increase the likelihood of reaching a successful resolution for the service failure according to Hart, Heskett and Sasser (1990). Ideally, the problem has to be solved before the customer is aware of the problem.
10
(apology). Some of these strategies can be divided into substrategies. Denial can be split in simple denial or shifting the blame on someone else. Evasion of responsibility can be divided into four strategies: provocation, defeasibility, accident and good intentions. Reducing offensiveness of event comprises bolstering, minimization, differentiation, transcendence, attack accuser and compensation. Furthermore, Benoit (1997) argues that it is important to understand the nature of the crisis, the perceived severity of the offence and to identify the salient audience.
Other service recovery strategies are mentioned by Huang (2008) in his research on the extent to which crisis communicative strategy and form of crisis response affect trust and relational commitment with respect to crisis contexts at the firm level. He shows three different responses: timely response, consistent response and active response. ‘Timely response’ becomes more critical in a situation when the external environment or state of an event is getting more unpredictable (Sillince 2002). Strong, Ringer, and Taylor (2001) also found in their study that timeliness of communication is one of the critical influencing factors of satisfaction across stakeholder groups and timely communication also fosters trust (Moorman, Deshpande and Zaltman 1993). ‘Consistent response’ is when information is being presented in a contradiction-free manner, which enhances credibility and accountability (Huang, 2008). The third strategy ‘active response’ means that the organization actively makes responses during the course of a crisis which is important since inactiveness or passiveness gives the appearance of indifference, uncontrollability, or hiding information (Huang 2008). Lastly, research by Coombs (2000) identifies seven different recovery strategies. Most of them are already covered by the previous mentioned strategies. The ones that are not clearly mentioned yet are attack-the-accuser (confront the one that claims a crisis exists), and ingratiation (seeking public approval).
2.2.2 Different media channels
11
Jin (2011) social networks and blogs account for one in every four and half minutes spent online worldwide. They also indicate that during crises publics use the internet even more. Moreover, people think that social media have a higher credibility level in some cases compared to traditional mass media, that they provide emotional support after crises and that they are a way to group virtually together and share (Jin, Liu and Austin 2011). Despite the benefits of social media, not all companies have fully implemented social media in their strategic communication (Liu, Austin and Jin 2011).
2.2.3 Five service recovery variables
Based on the discussed literature five service recovery variables are developed. These all influence the relation between a service failure and customer satisfaction. These variables are now explained.
Atonement The first service recovery variable in this thesis that is derived from the literature is
‘Atonement’. Research by for example Miller, Craighead and Karwan (2000), Levesque and McDougall (2000), Coombs and Holladay (2008), Tax, Brown and Chandrashekaran (1998) and many others indicate a clear relation between forms of atonement and customer satisfaction. One definition of atonement can be a repair done for the sake of a damaged relationship. Boshoff (2005) uses this term also for his RECOVSAT instrument. Atonement is present in multiple forms, these could be compensation, apology or a gift. The question to ask is which type of atonement consumers prefer when a company is facing a service crisis.
Content of communication The second service recovery variable that can be derived from the
literature is the content of communication. Multiple recovery strategies indicate that companies can respond in several ways in time of a service crisis (Benoit 1997; Coombs and Holladay 2008). Crises are always characterized by high levels of uncertainty, therefore information is considerably helpful in such situations (Stephens and Malone 2009). This uncertainty is generated by public’s crisis perceptions, organizational response, and issues concerning blame and cause. It can even be higher when the crisis involves complex technical details (Stephens and Malone 2009). Prior research has stated that organizations play an important role in providing their stakeholders with information (Lyon and Cameron 2004). In addition, Stephens and Malone (2009) state that when communication increases, people perceive the crisis to be less serious. Moreover, when customers face a service crisis they would like to have more information, for example about the problem and what causes it. Stephens and Malone (2009) state that when an organization is not providing any information, the customer will search actively on its own, for example online.
Communication media The next response recovery variable that can be derived from the literature is
12
can make use of several media channels. These can be the more traditional channels like television, Teletext and radio, but also more newer forms like news websites. Today, social media can also be used as a crisis communication channel and is growing rapidly (Liu, Austin and Jin 2011). Examples are Facebook, Twitter and social blogs. The question to ask here is which type of communication media used by the company is preferred by consumers when that company is facing a service failure.
Communication response time The fourth response recovery variable that can be derived from the
literature is the communication response time by the company that faces a service crisis. In this study, it is defined as the time elapsed from the moment the service crisis started till the moment the company sends out a communication message to its customers. A timely response can be very important for a customer since he wants to know what is going on. The more uncertain the service crisis, the more critical a timely response is (Sillince 2002). Moreover, as stated before, a study by Strong, Ringer, and Taylor’s (2001) indicates that timeliness of communication is one of the factors critical to satisfaction across stakeholder groups. Research regarding timing strategies in crisis communication suggests that the later communication is initiated in the crisis cycle, the less persuasive it will be (Arpan and Roskos-Ewoldsen 2005). However, it can be difficult for a company to release information quickly despite that a quickly crisis response seems both ethical and prudent. Many factors like fear of legal liability, the need to assess the situation, and the need to develop a unified organizational message are the cause of this (Arpan and Roskos-Ewoldsen 2005).
Solution response time The last response recovery variable that can be derived from the literature is
the solution response time by the company that is facing a service crisis. Important to note, this is not the same variable as the communication response time variable. In this study, it is defined as the time elapsed from the moment the service crisis started till the moment the company in question provides a solid solution. Acting fast is key here since customers want to use the service from the company as soon as possible and a slow recovery response by the company can deteriorate customers’ evaluations (Smith, Bolton and Wagner 1999). Fixing the problem is especially important in case of core failures since it is argued that service firms in that case have little leeway (Parasuraman, Berry, and Zeithaml, 1991).
2.2.4 Socio-demographic variables
13
written about age, however Chung-Herrera, Gonzalez and Hoffman (2010) indicate that equity perceptions in a service recovery was influenced by the age of the respondent. They found that younger groups are more easily to please compared to older age groups in terms of a recovery from a service failure . The gender variable is a second traditional segmentation variable which is easy to identify and record. According to literature, gender may be of relevance in understanding how service failure and recovery processes are perceived and processed (Palmer, Beggs and Keown-McMullan 2000). A few relevant studies exist in terms of the role of gender in service failures and recovery (Chung-Herrera, Gonzalez and Hoffman 2010). McColl-Kennedy, Daus and Sparks (2003) found in their study significant differences between males and females regarding their perceptions of service recovery handling. It appears that women want their views heard during service recovery attempts and want to be allowed to provide input, while men do not view voice as important. Another study by Iacobucci and Ostrom (1993) revealed that women are more demanding than men in a situation of a female service provider; women expect social skills on top of basic core service. Another variable that is used in this study is how many times a customer has experienced a service failure in a particular industry before in his life. It is to be expected that if customers have perceived multiple service crises before in that same industry that they are somewhat more experienced, in the sense that they already discovered the relevant and important elements of a service recovery. In addition, it can be expected that they are less quickly dissatisfied since they formed a general norm of service recoveries by companies. This expectation is based on Cadotte et al. (1987) who propose that previous experience with a given service is a primary determinant of expectations of service quality. A final variable relevant to the last one is the extent to which a customer was satisfied with the solution offered by the company that faced service failure.
2.3 Conceptual Model
To summarize and conclude this chapter a conceptual model is developed in
14
Service Recovery
Figure 1: Conceptual model
15
3. RESEARCH DESIGN
3.1 Research method: CBC analysis
To find out what the best response strategy is for a company to result in the best outcome of a major service crisis, it is useful to analyse what people’s preferences are. A good method for this is conjoint analysis, which is a multivariate technique developed specifically to understand how respondents develop preferences for any type of object (products, services, or ideas) (Hair et al. 2010). Multiple basic conjoint analysis methodologies exist of which the choice revolves around the basic characteristics of the proposed research: number of attributes handled, level of analysis, choice task, and the permitted model form (Hair et al. 2010; Orme 2003).
For this study the choice-based conjoint method (CBC) is chosen. This method employs a unique form of presenting profiles in sets; respondents choose one profile from a set of profiles rather than one by one (Hair et al. 2010). Since the task is more complicated, the number of factors included is limited, but the approach does allow for inclusion of interactions and can be estimated at the aggregate or individual level (Hair et al. 2010). Since this study uses five attributes, CBC is suitable. Another reason to choose for CBC is that it is argued to be more realistic, as it is much more representative of the actual process of selecting an object (or strategy in this study) from a set of competing options. Moreover, CBC provides an option of not choosing any of the presented profiles by including a non-option in the choice set (Hair et al. 2010). By selecting that non-option, a respondent can contribute information about the decrease in demand to be expected if all choices are unattractive (Sawtooth, 2013a). According to Johnson and Orme (1996), it is more realistic since in real life that option is also available and respondents are not forced to choose an unacceptable alternative.
After each respondent has chosen a profile for each choice set, the part-worths for each attribute level can be estimated. From these results, the contributions of each factor and factor-level interaction can be assessed and the likely shares of competing profiles can be estimated (Hair et al. 2010).
3.2 Study design of CBC
3.2.1 Number of attributes and levels
16
and most CBC studies use only three or four attributes. However, best results are achieved when keepings things simple (Johnson and Orme 2003). Orme (2002) indicates that more than six attributes in CBC can be difficult for respondents, in terms of boredom or fatigue. In addition, the levels are held equal to eliminate the ‘number of levels effect’ as far as possible, since if an attribute contains more levels than the other attributes, this attribute achieves a much higher average importance than the attributes containing less levels (Hair et al. 2010; Wittink et al. 1992).
In Table 2 can be seen that the first attribute ‘atonement’ consists of the levels compensation, gift and apology. ‘Compensation’ was in the survey further described as getting part of your money back. A compensation or refund strategy is according to Kelley, Hoffman and Davis (1993) a widely reported recovery strategy. Sparks and McColl-Kennedy (2001) found out that respondents expressed higher satisfaction with the service when a 50% refund was given. The ‘gift’ level is somewhat similar, but it differs in that the customer receive a present instead of part of the money back. In the survey, this gift was explained as extra minutes or data. The third level ‘apology’ is also a widely reported recovery strategy by for example Boshoff (1998), Coombs and Holladay (2008), Goodwin and Ross (1992), Kelley, Hoffman and Davis (1993) and Tax (1998).
Table 2: Attributes and their levels
Attribute Level 1 Level 2 Level 3
Atonement Compensation Gift Apology
Content of communication No information (only announcing existence of problem) Limited information (announcing what is the problem) Full information
(announcing what is the problem + its cause) Communication media Traditional media
(Tv, Teletekst and radio)
New media (news websites)
Social media
(Facebook, Twitter, blogs) Communication response
time
< 1 hour 2-4 hours 5-24 hours
Solution response time 1-4 hours 2-3 days At least 1 week
17
The third attribute ‘communication media’ consists of the levels traditional media, new media and social media. ‘Traditional media’ comprise the following channels: television, Teletekst and radio. Chosen is not to include printed media since, as mentioned before, more people receive their news from television than from print (Coombs and Holladay 2009). Moreover, printed media (e.g. newspapers) are relatively ‘slow’; printing (and delivering) takes time. However, in case of a service crisis fast communication is required. The ‘new media’ level is defined as news websites. This category is chosen since news websites do not belong to either ‘traditional media’ or ‘social media’ channels. ‘Social media’ is further defined as Facebook, Twitter and blogs.
The fourth attribute ‘communication response time’ consists of the ordinal levels < 1 hour, 2-4 hours and 5-24 hours. All these levels are chosen not to be longer than one day, since nowadays it can be assumed that the norm for a company is to response in a day, otherwise it does not take its customers seriously. The final attribute ‘solution response time’ consists of the ordinal levels 1-4 hours, 2-3 days and ‘at least one week’. This distribution is based on examples from practice where failures sometimes are solved in one day and in other cases after weeks.
3.2.2 Design of choice task
The choice task in this study uses full-profiles which is normal in a CBC study. This means that the alternatives contain all five attributes. A fractional factorial design is chosen to reduce the number of profiles and choice tasks presented to respondents. This method is most common for defining a subset of profiles for evaluation (Hair et al. 2010). Since five attributes with each three levels make 249 different profiles possible which is impractical. Therefore, 41 profiles are generated by the Sawtooth program (excluding the hold-out task, which will be explained later).
18
3.2.3 Holdout task
Both survey versions include the same fixed holdout task in this study. These tasks typically look just like regular CBC choice tasks, but they are not used (‘held out’) during utility estimation (Johnson and Orme 2010). They are useful for multiple reasons, one is to check validity of the estimated utilities, measured by the utilities’ ability to predict choices not used in their estimation (Johnson and Orme 2010). In designing hold out tasks, it is important that the profiles are not equally attractive or that only one of the choices is preferred (Johnson and Orme 2010). They indicate that the position of the holdout set is important in the survey since it is well known that the first CBC tasks contain the most noise and the lowest scale and respondents tend to learn through the process of completing multiple CBC tasks. Sawtooth has placed the holdout set at the third place of the total eight questions in the survey. The holdout set can be found in Appendix B.
3.3 Experimental procedure
The data is collected by making use of a survey. The conjoint questions asked in this survey are generated by the software program Sawtooth SSI Web version 6.6.6. These questions and answers had to be translated into Dutch since the survey would be distributed among Dutch individuals. The survey could be accessed by making use of an online survey software tool called SurveyGizmo (www.surveygizmo.com) so that people only had to click on the generated web link. This link was distributed to as many people as possible (family, friends and others, via Facebook and email). Snowball sampling is applied by asking to share the link with others. Before the survey went online, SurveyGizmo provided a small diagnostic check which tests the estimated length, complexity, fatigue score and accessibility. The survey passed the test, all checks were good (Appendix C).
The survey began with an introduction (Appendix D), followed by simple demographic questions and some other subject related questions. Next, a screen was presented providing detailed information about the coming conjoint questions and procedure. A table with the attributes and levels was provided to make it more understandable for respondents. It was made clear that they had to choose one of the four presented options by clicking on them. Basically, the first three options can be seen as three different ‘strategies’ which can be applied by a mobile telephone provider to deal with a failure. The last option was the none-option, which could be selected if none of the three options appealed to the respondent.
19
perceived that attribute by making use of a slider (ranging from 0-100). This information can be used for a reliability check. In the final screen the respondents were thanked for their participation.
3.4. Plan of analysis
The analysis starts with a sample description to get deeper insight in the respondents’ data. Next, the data is analysed at aggregated level, which is a traditional method for a choice based conjoint study. Latent class modeling in Latent Gold Choice software is used to estimate the part-worth utilities of the respondents. Firstly, the distribution of the attributes is examined. Multiple model specifications exist; these are also called preference functions. Vector (linear) models are the simplest since less parameters have to be estimated compared to the other models. More complex is an ideal point (quadratic) model and most complex is a part-worth model (Hair et al. 2010). It is important to note that it is not required that all attributes have the same type of model. It is to be expected that none of the attributes in this study can be estimated with another model except for a part-worth model, in other words no linear relationship is expected. When all attributes and the non-option are set to nominal (part-worth) and effects coding is used, the estimated parameters for all attributes can be plotted to check the linearity. The best model is chosen based on the often used model performance indicators Bayesian information criterion (BIC), Akaike information criteria 3 (AIC3), consistent Akaike information criterion (CAIC) and the weight of evidence AWE. For interpretation purposes, the parameter estimates and the relevant importance are analysed. The predictive validity is checked by calculating the in-sample and out-of-sample hit rates.
20
weighted averages of the class-specific effects, where the posterior membership probabilities of a case serve as weights (Vermunt and Magidson 2005).
21
4. ANALYSIS AND RESULTS
4.1 Sample description
In total, the link of the survey was clicked on 337 times, of which 203 respondents (60.2%) finished the survey. The other 134 respondents (39.8%) abandoned the survey. Some left right after clicking on the link or somewhere during the survey, but most of them left at the beginning of the conjoint part. However, not all 203 respondents can be used for latent class analysis, since four respondents chose the non-option for all conjoint questions. These respondents can be seen as outliers and are deleted as they do not provide relative information about the attributes they prefer. Therefore, in total 199 respondents are used for further analysis in this study. Five respondents in this group finished the survey above one hour, one of these even needed more than eight hours. A logical explanation for these ‘outliers’ is that they were not continuously busy with filling in the survey, so that they continued the survey at a later point in time. It can be argued that this is too long, however decided is to keep these five respondents in the data since they provide valuable data. The average completion time without these respondents is around nine minutes, where most respondents completed the survey in seven minutes.
22
Table 3: Sample description
Variable Categories % Variable Categories %
Age 15 – 24 25 – 34 35 – 44 45 – 54 55 – 64 65+ 25.1 6.1 8.0 22.1 18.6 20.1 Experienced failures before never 1 – 3 times 4 – 6 times > 6 times 15.6 59.8 12.6 12.1 Gender Education level male female elementary school mavo, vmbo, mbo 1 havo, vwo, mbo 2-4 hbo wo 60.8 39.2 2.0 15.6 20.1 35.2 27.1 Satisfied with solution Very dissatisfied Dissatisfied Moderately dissatisfied Neutral Moderately satisfied Satisfied Very satisfied Not applicable 1.0 8.0 9.5 20.1 27.6 18.6 0.5 14.6
4.2 Latent class analysis – Aggregate level
Latent class modeling in Latent Gold Choice software is used to estimate the part-worth utilities of the respondents. Firstly, the distribution of the attributes is examined. As mentioned before, it is to be expected that none of the attributes in this study can be estimated with another model except for a part-worth model. In other words, no linear relationship is expected. When all attributes and the non-option are set to nominal (part-worth) and effects coding is used, the estimated parameters for all attributes can be plotted to check the linearity. None of the graphs are linear, however the linearity of ‘solution response time’ is arguable (Appendix E). This is probable because in general everyone prefers speed over slow in providing a solution to a problem. Therefore, the model is also estimated with this attribute set to numeric which leads to fewer parameters in the model. However, the general model performance of this model is worse compared to the first model, since the often used model performance indicators Bayesian information criterion (BIC), Akaike information criteria 3 (AIC3), Consistent Akaike information criterion (CAIC) are slightly higher (worse). Only the approximate weight of evidence AWE is lower which is better. Therefore, it is decided to use the model in which all attributes are estimated with the part-worth model. In Table 4 the model fit statistics are given.
Table 4: Model fit statistics
Model BIC AIC3 CAIC AWE R2
23
This model is an aggregated model, which means that all respondents are in one group. In addition, this model is significant and explains 12.64% of the variance in the dependent choice variable. Furthermore, all attributes including the non-option are significant. This implies that different levels within an attribute have significantly different effects on the utility relative to the overall average utility.
Interpretation When looking at the part-worth utility estimates of the attributes in Table 5, the
preferences of the respondents are known. Starting with the atonement attribute, it turns out that a compensation is most preferred followed by a gift. However, an apology is not preferred at all since this utility estimate is negative. Next, respondents only prefer full information for the communication content attribute, since ‘no information’ and ‘limited information’ have negative utility estimates. For the communication media used, it is noteworthy that respondents value the traditional channel (Tv, Teletekst and radio) most, however, almost the same as new media (news websites). Striking is that the social media channel is not valued at all, since it has a negative utility estimate. For both the communication response time and the solution response time the preferences are pretty the same in the sense that they prefer faster over slower levels. However, the difference is that only the fastest level of 1-4 hours is valued for solution response time while for communication response time only the slowest level of 5-24 hours has a negative estimate. The non-option part-worth utility estimates indicate that choosing the non-option is not preferred since the estimate is negative. This means that respondents prefer to choose one the three options with the attributes.
Table 5: Parameter estimates for the aggregated model
Attribute Part-worth
utility estimate
P- value Attribute Part-worth utility estimate
P- value
Atonement 1,8e-7 Communication
response time 0,0089 Compensation 0,1626 < 1 hour 0,1019 Gift 0,1131 2 – 4 hours 0,0406 Apology -0,2757 5 – 24 hours -0,1424 Communication content
5,3e-13 Solution response time
1,1e-69
No information -0,2332 1 – 4 hours 0,7232
Limited information -0,0871 2 – 3 days -0,1384
Full information 0,3202 At least 1 week -0,5848
Communication media 0,0025 None option 3,7e-15
Traditional media 0,0869 No 0,3291
New media 0,0780 Yes -0,3291
24
Relative importance The part-worth’s have no exact meaning because they are estimated on an
interval scale with an arbitrary origin. This has also the consequence that levels of one attribute cannot be compared to other attribute levels (Hair et al. 2010; Orme 2010). Therefore, the relative importance of the attributes is estimated since it is interesting to know which attributes the respondents relatively find important. The attribute that is found most important is the ‘solution response time’ (37.9%) (Figure 2). After this, the non-option is most important (19.1%). Next, the communication content is most important (16.0%) followed by atonement (12.7%). Communication media and communication response time are less important for the respondents with respectively 7.3% and 7.1%.
Figure 2: Relative importance of attributes (aggregated model)
Predictive validity From the prediction statistics provided by Latent Gold Choice software the
internal validity can be calculated. This is also called the in-sample hit rate and is based on the mode (Vermunt and Magidson 2005). A hit rate means the extent to which the model is correct in predicting the choices that are made by the respondents, which is 52.3% for this aggregated model
(
Table 6
). Noteworthy is that the model fails to predict all 172 observed non-options correctly. Theout-of-sample hit rate can also be calculated by making use of the holdout set discussed before. If the model accurately predicts the chosen alternative in the hold out task, a ‘hit’ is observed. The model has an out-of-sample hit rate of 40.7%. Compared to random prediction which would lead to a hit rate of 25%, this model performs better. When this hit rate is benchmarked, it is a little below some other cbc hit rates on the aggregated level. For example, Vriens, Oppewal and Wedel (1998) have a hit rate of 52% and Moore, Gray-Lee and Louviere (1998) a hit rate of 43%.
Table 6: Prediction table for the aggregated model
25
4.3 Latent Class analysis - Segment level
Instead of estimating the model at aggregated level, a model can also be estimated at segment level. An advantage of segment level models is that they take the heterogeneity in consumer preferences into account by dividing the respondents into different classes. Latent Gold software can divide them into latent classes with similar preferences in the groups (homogeneity) and different preferences between the groups (heterogeneity), based on the choices made in the CBC questions (Sawtooth 2004). Latent Gold software also offers the ability to predict class membership not only based on preferences but also on individual characteristics (covariates) which gives the ability to predict to which (unobserved) subgroup an individual belongs (Vermunt and Magidson 2005). In the program, the covariates can either be set to active or inactive, which will lead to different models.
Model fit statistics To determine which model to use and the number of segments, multiple models
with different segment numbers and sizes are compared. First, models were generated with all covariates set to active which is indicated in Table 7. In this situation, the three segment model is the best model according to the information criteria BIC and CAIC. However, taking a closer look to the covariates reveals that only education is significant for predicting a 4-class solution (Appendix F). Moreover, the attribute ‘communication response time’ is only significant in a 4-class solution.
Table 7: Model fit statistics all covariates set to active
Model BIC AIC3 CAIC AWE R2
2 segments 3173.23 3083.79 3212.23 3522.88 0.2219
3 segments 3136.15 2982.50 3203.15 3752.48 0.3298
4 segments 3177.95 2960.08 3272.95 4018.80 0.3748
5 segments 3245.56 2963.48 3368.56 4316.91 0.3989
To see if better models could be estimated, many estimations are performed by varying the covariate composition (active/inactive). The covariates ‘gender’, ‘experienced failure before’ and ‘satisfied with solutions’ are never significant for predicting class membership, so these will be set to inactive. The covariate ‘age’ is only significant for a two and five segment solution. The covariate education appears to be only significant for a four and five segment solution. When both ‘age’ and ‘education’ are set to active, ‘age’ is never significant and ‘education’ only with a four segment solution. However, the information criteria of this model are worse compared to the four segment solution with only ‘education’.
26
‘education’ as only active covariate will be a good choice when taking the information criteria and the more practical four segments into consideration. Taking a closer look to the parameters indicates that the communication response time attribute is still not significant (p = 0.081) at the 0.05 level and not significant for the test of equality (p = 0.170). Therefore, it is decided to restrict the effect of this attribute to be equal across all four segments. By setting the ‘Class Independent’ to ‘Yes’ in Latent Gold Choice for this attribute, the p-value for communication response time becomes significant (p = 0.030). Furthermore, the model fit increases (see last model in Table 8).
Table 8: Model fit statistics - Different covariates
Model BIC AIC3 CAIC AWE R2
2 segments AGE 3102.97 3047.93 3126.97 3327.37 0.2221
5 segments AGE 3008.69 2864.21 3071.69 3593.58 0.4196
4 segments EDU 3032.21 2896.90 3091.21 3577.95 0.3714
5 segments EDU 3046.17 2874.17 3121.17 3726.00 0.4192
4 segments EDU & AGE 3045.00 2902.82 3107.00 3617.44 0.3728 4 segments EDU (resp time
equal for all segments)
3007.03 2885.48 3060.03 3509.27 0.3762
EDU = Education resp time = communication response time attribute
Table 9: Most attribute levels per segment
Segment Atonement Communication content Communi-cation media Communi-cation response time Solution response time None-option
1 (43.3%) Gift Full information Traditional < 1 hour 1-4 hours No 2 (31.0) Gift Full information New < 1 hour 1-4 hours No 3 (17.6%) Compensation Full information Traditional < 1 hour 1-4 hours Yes 4 (8.1%) Apology Limited
information
Social < 1 hour 1-4 hours No
Interpretation The four obtained segments have different sizes. The first segment includes 43.3% of
27
value a gift most, however segment 1 also values a compensation highly. The third segment values the same communication media as segment 1, however they prefer a compensation instead of a gift. The smallest segment prefers an apology and values social media most as communication media. Another difference that can be found is that segment 3 is the only segment that values choosing none of the presented ‘strategies’ which exist of these recovery attributes. A complete overview of these parameter estimates is given in Appendix G.
Figure 3: Relative importance of the attributes per segment
To provide more insight, the relative importance of the attributes is analysed (Figure 3). Segment 1 has the highest relative importance for the non-option (48%). This means in this case that this segment always tries to choose one of the three options and not the non-option since that parameter is very negative (see Table 9). The second most important attribute is atonement (18%). For segment 2, solution response time is by far the most important attribute (52%) followed by communication media (16%). Segment 3 also finds solution response time most important, however, its share (33%) is lower compared to segment 2. The second most important attribute is the non-option (29%). However, in this case it means that this segment tries to choose the non-non-option instead of one of the three presented options since this is preferred (Table 9). The last segment finds atonement most important (38%) followed by solution response time (25%). Furthermore, it can be
28
seen that response time solution is least preferred by all segments and communication content is around 11% or 14% for all segments.
Differences in consumer characteristics The respondents in segment 1 have the same gender
distribution as the whole sample, which is logical since this is the largest segment (Appendix H). The other segments differ slightly from this distribution. Segment 2 consists of the youngest respondents as the average age is lowest (42 years old) due to the fact that 28% of this segment is between 15 and 23 years old. Segment 3 has the highest average age with 52 years old. Segment 1 has less respondents with a hbo education level compared to the overall distribution (21.5% vs. 35.2%) which is confirmed by the very negative corresponding parameter (Appendix G), but has a higher havo, vwo mbo2-4 educational level (29.7% vs. 20.1%). This is also confirmed by the very positive corresponding parameter. In addition, striking is that segment 2 is relatively highly educated as it consists of 48.7% hbo and 34.5% wo respondents which are both above the averages. Again, the parameters for these covariate levels are most positive. Segment 3 has also a large hbo group (45.7%), but a smaller wo group (18.6%), which is also indicated by the corresponding negative parameter. Segment 4 has many respondents that have a mavo, vmbo, mbo 1 educational level (confirmed by the very positive corresponding parameter) and compared to the other segments the lowest havo, vwo mbo2-4 educational level (6.3%), the lowest wo level of all (13.4%) and the lowest elementary school level (0.0%). This is confirmed by al the corresponding negative parameters. The differences in experienced a service failure before by the respondents are small. Notable is that segment 1 has the lowest percentage of people who never experienced such a failure (11.4%) and that segment 2 and 4 have the highest percentages of people for this covariate (19.7% and 19.5%). The satisfaction level covariate has minor differences, it indicates that 11.7% of segment 4 consists of very dissatisfied customers while other segments have 0.0% for this category. However, they are a lot more ‘dissatisfied’ and ‘moderately dissatisfied’ while segment 4 has very few of these categories. Lastly, segment 1 had the most satisfied respondents with 28.7% while the other segments are around 10-12%.
Predictive validity From the prediction statistics the internal validity can be calculated again. This is
29
this study are both respectively larger compared to the hit rates 50.8% and 40.7% of the aggregated model, as could be expected.
4.4 Reliability check
An alternative method to analyse if certain respondents value different attributes, is ordinary linear regression (OLS). In Table 10 the parameter estimates can be found for all five models. The notable findings are that gender (Female) and experienced failure before are never significant in explaining the attribute ratings. In contrast, the education level is significant for explaining the importance of the communication content, communication media and the communication response time attribute. In addition, the parameters are all negative which means that a higher education level results in a lower importance of communication content, communication media and communication response time. The covariate age is only significant for explaining the atonement attribute rating and has a negative parameter which means that when age becomes higher the atonement attribute rating decreases. The satisfaction level covariate is only significant for explaining the communication media attribute rating and has a positive parameter. This means that a higher satisfaction level of the solution leads to an increase in the importance of communication media.
Table 10: Regression parameter estimates
Atonement Communication content Communication media Communication response time Solution response time Constant 71.082* 80.602* 38.188* 84.317* 94.540* Female+ 7.030 3.290 2.262 2.144 3.538 Age -0.342* -0.056 0.171 0.011 -0.065 Education level -1.951 -5.647* -4.907** -4.790* 0.163 Experienced failure before -3.668 -2.007 -1.415 -1.033 -0.989 Satisfaction level 0.836 0.874 4.188* 0.882 -2.131
* = sign. at 0.01 level ** = sign. at 0.05 level + = Male is set to zero
30
5. CONCLUSION
This study examines which action and communication response strategies are best to implement by service companies that face a major crisis/failure. Research on major general service crises is limited since most literature is on the individual service encounter level, despite the major impact. Firms that are facing such major crises often react in a rather clumsy way to deal with the problem. In addition, research has been done on recovery strategies, however, most relevant studies do not focus on the importance of the different service recovery attributes. These attributes have a moderating effect on the negative relation that a service failure has on customer satisfaction. The contribution of this study is that a general theoretical framework is built based on relevant literature to research which of these recovery attributes are found to be most important by customers. In this study, a mobile telecom provider is taken as an example for a service company which experiences a major service crisis (e.g. a major connectivity failure). Data is collected by means of an online survey distributed under family, friends, students or other known acquaintances with applying snowball sampling. Choice based conjoint analysis is used for identifying the preferences of respondents, which are analysed with latent class analysis. Since the nature of this study is somewhat explorative, no real expectations are formulated.
5.1 Discussion results
31
respondents value compensation (part of money back) more than a gift (extra minutes or data) as a form of atonement, and an apology is not valued. This is in line with research by Coombs and Holladay (2008) who state that an apology is not the best recovery strategy. In this study, respondents could only choose profiles with one level per attribute. It could be that when combinations of multiple atonement levels were possible, results were different since Goodwin and Ross (1992) indicate that when no tangible offering is made, apology has a lower effect. Communication media and communication response time are found to be least important. Striking however, is that social media are not valued as a communication medium despite its rapidly growing importance and its higher credibility level in some cases, compared to traditional mass media (Liu, Austin and Jin 2011). A logical explanation for this can be that around half of the respondents are over 50 years of age. According to Pew Internet & American Life (2013a), only half of the 50-64 age group and only 32% of the 65+ age group use social networking. Another striking finding is that communication response time has really low relative importance, as it can be frustrating to wait for an explanation why a service is not working for example.
32
Segment 1 has a high share (30%) of an average education level (havo, vwo and mbo2-4), which is significantly higher compared to the other segments. Also, only 11% of respondents in this segment experienced a service failure before.
Segment 2 has the highest relative importance for solution time (52%), they really value a failure that is solved as soon as possible (1-4 hours). Explanation for this can be that this segment has the lowest average age. This group is very dependent on mobile connectivity and are very active users (Pew Internet & American Life 2011). Striking is that segment 2 is highly educated since more than 80% has an education level of hbo and wo, compared to 50% for segment 1.
Segment 3 has most similarities with segment 1, but segment 3 values only a compensation as atonement while segment 1 values a gift most, followed by a compensation. The highest relative importance for segment 3 is solution response time.
Segment 4, which is the smallest and the only segment that values social media most as communication media. Notable is that this segment has a relatively high share of a lower education level (mavo, vmbo, mbo 1). Furthermore, it has the highest relevant importance for atonement and is the only segment that values an apology most. Lastly, it is striking that for the solution response time attribute (which has second highest relative importance) level ‘2-3 days’ is not valued, however ‘at least 1 week’ is slightly positive valued. As mentioned before, these levels are ordinal so ordinal preferences could be expected. No logical explanation can be given for this outcome.
5.2 Managerial recommendations
33
Holladay 2008), a possible suggestion is to apologize in combination with either a compensation or a gift. Despite being one of the two least relatively important service recovery attributes, it is recommended for the communication strategy to use traditional media channels (Tv, Teletekst and radio) and new media (news websites) to reach and inform customers. According the results, social media like Facebook, Twitter and blogs were not valued for communication purposes. However, since its rapidly growing importance it is recommended to already explore the possibilities (Liu, Austin and Jin 2011). The other attribute that is least relatively important is communication response time. It is recommended for the communication strategy to inform customers as soon as possible about the service crisis, preferably within an hour after the start, but within 2-4 hours will also work. However, a slower response is not recommended.
This study could identify four different groups of customers. If a service company has the opportunity and possibility, it could try to implement the right recovery strategies which fit best for their customer base to achieve high satisfaction for all customers. The customers in segment 1 will probably be most content with a recovery strategy which exist of attribute levels they valued most in this study. Therefore, the focus for the action strategy should be on a gift (or compensation) and for the communication strategy giving full information about the service failure. Despite being less important, the solution response time should be fast (1-4 hours), the communication response time should be fast (< 1 hour) and traditional media (Tv, Teletekst and radio) should be used. This segment is important since it size is 43% and it can be identified by an average education level (havo, vwo, mbo 2-4) and by the fact that most of the consumers have experienced multiple service failures before.
For the second largest segment, which includes relatively the youngest individuals and individuals with a higher education level (hbo and wo), a somewhat different strategy should be followed. Here, the focus should especially be on an action strategy that fixes the service failure as soon as possible, since the other attributes have a significantly lower relatively importance. Despite the lower importance, the communication strategy should consist of reaching the customers as soon as possible by means of news internet sites providing full information about the service crisis.
