Click or Call to Predict Customer Switch:

(1)

Click or Call to Predict Customer Switch:

discovering the predictive value of clickstream and service call

center data on customer switch in a service company setting.

(2)

Click or Call to Predict Customer Switch:

discovering the predictive value of clickstream and service call center data on

customer switch in a service company setting.

Liza Willemsen University of Groningen Faculty of Economics and Business

Msc Marketing Intelligence and MSc Marketing Management Master Thesis June 20, 2016 Mathenesserdijk 101A02 3027 BE Rotterdam Student number: s1964011 Email: l-willemsen@live.nl Tel: +31 (0)619053339 Supervisors University of Groningen

First supervisor: prof. dr. J.E. Wieringa (j.e.wieringa@rug.nl) Second supervisor: dr. J.T. Bouma (j.t.bouma@rug.nl)

University of Groningen Faculty of Economic & Business

Department of Marketing

(3)

Executive Summary

Since the liberalization of the energy market in 2004, customer retention has become a major issue for all energy providers. Where no customer retention initiatives were needed in the past, energy providers are now fighting to prevent their customers from switching.

Until today some knowledge about the determinants of customer switch in the energy industry was available, though was based on traditional marketing measures like relationship perceptions, marketing instruments and customer characteristics. However, not much is known about the relatively new data provided by modern communication and service channels like the website and service call center.

This study aims to find the predictive value of these channels. Therefore, a logit model was developed to find what online and service call center behavior does predict customer switch. The data used for this logit model was provided by a Dutch energy provider. In order to increase the predictive accuracy of the model, a balanced stratified sample was created. The model showed that, opposite from what was expected, the total amount of website visits had a negative effect on customer’s switching probability. The amount of total time spent on the website was found to increase the customer switching probability. On a more specific level it was found that those customer who visited in more than 50% of all their sessions a FGC page type, were more likely to switch than those customer that did not visit these pages at all. Moreover, in line with the expectations, prove was found that customers who visited pages containing information about the current product were more likely to switch than those customer that did not visit these pages at all. Furthermore, the results showed that those customers that visited pages containing information about additional products in up to 50% of all their visits, were less likely to switch than those customer that did not visit these pages at all. No significant effect was found for the service call behavior.

(4)

call. However, the advantage the online data offers is that the customer behavior can be tracked real time, which creates opportunities for real-time interventions. It is recommended to build an automatic system that uses the model to send automated retention actions during the browsing session of the customers. These actions should have a customer oriented focus and therefore it is recommended to create a service dialogue. Next, it also provides inspiration to smartly make use of the customer online account by personalizing the website. Lastly, the results emphasize the importance of stimulation of customer use of online accounts and therefore it is recommended to marketing managers to think of ways to stimulate this.

(5)

Preface

During my MSc Marketing I developed a high interest in customer loyalty, especially in ways to increase customer loyalty and retain valuable customers. However, I soon realized that companies do have to spend their marketing budgets carefully and do need handles and prove to base their budget allocation decisions on. This is how my interest in predicting models grew: these models can really make a difference for managers. Together with my personal interest in online marketing and websites, a thesis was born.

My interest alone was not enough to successfully execute this research. First I would like to thank my supervisors prof. dr. J.E. Wieringa and dr. J.T. Bouma for supervising my research by providing valuable feedback. Second I would like to thank my supervisors from Sunwinergy, Jan Jaan and Peter Peets for giving me the opportunity to discover marketing intelligence in practice at Sunwinergy and their time invested in guiding me during the internship period. Last I would like to thank my fellow students, friends and family for their unconditional support.

(6)

Inhoud

Executive Summary ... 2

Preface ... 4

Chapter 1: Introduction ... 8

Chapter 2: Theoretical Framework ... 10

2.1 Customer switch ... 11

2.2 Online customer behavior ... 13

Clickstream data ... 13

Website quality ... 17

Type of page visited ... 17

2.3 Customer service call behavior ... 21

Service call center ... 21

2.3.1 Incoming service call ... 21

2.4 Control variables: customer characteristics and relational characteristics ... 22

2.4.1 Age ... 22 2.4.2 Type of contract ... 22 2.4.3 Usage of service ... 23 2.4.4 Additional product ... 23 2.4.5 Additional charges ... 23 2.4.6 Length of relationship ... 23 Chapter 3: Methodology ... 24 3.1 Data collection ... 24 3.1.1 Sunwinergy ... 24 3.1.2 Original Dataset ... 24 3.2 Data Cleaning ... 25

3.2.2 Outliers, oddities and missing values ... 26

(7)

3.3.1 Stratified random balanced sample set (switch vs non-switch) ... 27

3.4 Measurement approach ... 29

3.4.1 Total amount of total website visits ... 29

3.4.2 Total time spent on website visits ... 30

3.4.3 Path length ... 30

3.4.4 Total amount of error notifications ... 30

3.4.5 Type of page visited ... 30

3.4.6 Incoming service call ... 31

3.4.7 Control variables: customer characteristics ... 31

3.4.8 Customer switch ... 31 3.5 Research design ... 34 3.6 Model specification ... 34 Chapter 4: Results ... 36 4.1 Descriptive statistics ... 36 4.2 Model choice ... 36

4.2.1 Added predictive value of online customer behavior and customer service call behavior ... 38

4.3 Variable investigation and sample set ... 39

4.3.1 Pearson Correlation Coefficients ... 39

4.3.2 Multicollinearity ... 39

4.4.3 Training and validation sample ... 39

4.5 Evaluation of the estimated model ... 40

4.5.1 Parameter estimates ... 40

4.5.2 Parameter estimates interpretation ... 42

4.6 Predictive validity ... 42

4.6.1 Robustness check ... 43

(8)

5.1 Findings ... 47

5.1.1 Total online browsing behavior ... 48

5.1.2 Website quality ... 49

5.1.3 Type of page visited ... 49

5.2 Scientific implications ... 52

5.3 Managerial implications ... 53

Chapter 6: Limitations and Further Research ... 55

(9)

Chapter 1: Introduction

One of the markets in the Netherlands that has faced a major change in the past decennia is the energy market. For many years customers had no free choice in energy providers, as these were locally determined. However, a response of the Dutch government to two EU electricity directives changed the rules of the game: the market was completely liberalized at July 2004 (van Damme, 2005). By this, the opportunity for customers to switch arose and made the energy market an attractive market for competition to enter. Therefore, it is not surprisingly that since the liberalization of the market, the amount of switching customers from their incumbent energy provider towards another one has increased each year; the energy market’s switch rate was 6,5% in 2007 and increased to a percentage of 14% in 2015 (Autoriteit Consument en Markt, 2015). According to the study of Autoriteit Consument en Markt (2015), almost half of the consumers state to have switched to another energy provider at least once since the liberalization of the energy market. This trend is expected to continue (Grol 2016), especially because today’s customers have become more price sensitive due to the transparency the internet offers, e.g. by price comparison sites(Jung, Cho & Lee 2014). Moreover, consumers have become more critical towards services (American Express 2012), which makes it harder for energy providers to satisfy their customers. Consequently, retaining customers is no longer a matter of course today. Jones, Mothersbaugh and Beatty (2000) have shown that the amount of perceived alternatives have a negative effect on renewing contracts, and thus on customer retention.

(10)

In order to retain more customers, companies should know which factors determine whether a customer is likely to stay with the company, or in other words, to switch to a competitor. By predicting which customers are likely to switch, companies can proactively intervene this process in order to prevent these customers from switching which creates all the (monetary) benefits described above. Hence, predicting a customer’s probability to switch provides a huge potential revenue source for companies.

(11)

yet. Naturally, all other managerial implications found by this study will be presented to enrich practical knowledge.

This study aims to answer the following main research question:

What customer online and service call behavior predicts customer switch in a service company?

With the following supporting sub questions:

1. What types of online browsing behavior exist?

2. Is there a link between online browsing behavior and customer switch? 3. Is there a link between call center interaction and customer switch?

In order to analyze this, customer data of a major Dutch energy provider Sunwinergy is used. Online browsing behavior is measured for those customers who were logged into their online Sunwinergy account. Data containing their service call behavior was added to this.

The paper is structured in the following way. First an overview of current literature is given in order to elaborate on the topic of this study (chapter 2 ‘Theoretical Framework’). Within this chapter several hypotheses are formed, which form the basis of the research. Chapter 3 ‘Methodology’ presents the research design that is followed by the results found (chapter 4 ‘Results’). These results are discussed in chapter 5 ‘Discussion and Conclusion’ and based on this, managerial implications will be given. This paper is closed by discussing the limitations of this study and giving direction for further research (chapter 6 ‘Limitations and Further Research’).

Chapter 2: Theoretical Framework

(12)

2.1 Customer switch

Within the marketing and service literature customer switch is also referred to as customer churn. Customer churn in the service industry is the termination of a relationship, where the company has lost the customer (Risselada, Verhoef & Bijmolt 2010). This can be caused not only by switching to a competitor, but also by death or movement to another location. As this study focuses specifically on those customers switching to a competitor, the term customer switch will be used instead of customer churn. Thus, in this study, customer switch is defined as the termination of the relationship with the customer caused by the customer switching to a competitor. Customer switch receives much attention in the marketing literature. This is of no surprise, as establishing long-term relationships and maximizing customer value is the main focus of many companies (Risselada et al. 2010). To achieve this, customers must be retained in the company. However, this can be a great challenge in highly competitive markets where customers have the freedom to choose between many different providers.

The reasons for switching to another firm are diverse and have been studied by many scholars. Within these studies two type of antecedents are studied, namely customer relationship perceptions like customer satisfaction or commitment/trust and marketing instruments like advertising and loyalty programs (Verhoef, van Doorn & Dorotic 2007). An example of a study that studies customer relationship perceptions is the one of Keaveney (1995). Within this study she identified reasons for customer switching in many different type of services industries, covered in eight general categories: pricing, inconvenience, core service failures, service encounter failures, employee responses to service failures, attraction by competitors, ethical problems and involuntary switching and seldom-mentioned incidents. Bansal, Taylor and James (2005) distinguished between these factors by linking them to overarching effects namely push effects, mooring effects and pull effects. The push effects are the firm’s behavior that causes customers to switch. Mooring effects are the customer’s personal inhibitors and facilitators for switching (Bansal et al. 2005). The last type of effects found are the pull effects, these effects are caused by competition which pull customers away from the firm. In their study they found that the last two type of effects play a more important role in customer switching behavior than the push effects.

(13)

telecom industries (Verhoef 2003, Bolton, Kannan & Bramlett 2000, Gustafsson, Johnson & Roos 2005, Bolton 1998 ) but also the energy industries (Walsh et al. 2005, Hartmann & Ibáñez 2007, Wieringa & Verhoef 2007).

Within the financial sector, Verhoef (2003) found that affective commitment and loyalty programs are significant determinants for customer retention. A similar finding is found by Bolton et al. (2000). In their study they found that those customers who are in a loyalty program “overlook or discount negative evaluations of the company vis-à-vis competition”. Hence, in their study it appears as a moderating effect.

Contrary to the findings of Verhoef (2003), Gustafsson et al. (2005) found that customer satisfaction predicts customer retention and not affective commitment. Next to this, they also found that those customers who have a history of prior switch are more likely to switch this time as well. Their study was applied in the telecom industry. Another study applied in the same industry is the one by Bolton (1998). In this study the author found that previous customer experiences with the company have an important influence on the positive effect of customer satisfaction on customer retention in the way that customers value previous experiences more heavily than new information.

Also in the energy market customer satisfaction appears to be an important factor predicting customer switch (Walsh et al. 2005). The study by Hartmann and Ibáñez (2007) confirms this, but also found that brand trust and switching costs are of equally important predictors of customer switch. Similar determinants were found by Wieringa and Verhoef (2007) in the Dutch market. According to their study, next to switching costs, also the relationship and current demand for products and services from the energy supplier are important determinants for customer switching behavior.

(14)

2.2 Online customer behavior

Clickstream data

In the past decades literature paid more and more attention to the richness of clickstream data. Bucklin and Sismeiro (2009) define clickstream data as “the electronic record of a user's activity on the Internet”. This is a record of the path a visitor takes within and across websites and is a reflection of the choices the visitor has made. Consequently, this type of data offers many opportunities for better understanding customer choice behavior (Bucklin et al. 2002) and has been used by scholars to pursue a wide range of topics. In their paper, Bucklin and Sismeiro (2009) group the studies into three broad research themes: (1) browsing and navigating behavior, (2) how it is used for advertising methods and (3) online shopping behavior. This study applies to the first and last research themes, therefore the second theme is not discussed more deeply.

Within the field of browsing and navigating behavior, the study of Moe (2003) defined four types of browsing behavior, namely directed buying, search and deliberation, hedonic browsing and knowledge building. The first two are goal-directed in nature while the last two are more exploratory. Other studies focus more on site navigation choices like the study of Bucklin et al. (2002) and Buckling and Sismeiro (2009). These include choices concerning within-site browsing (Bucklin & Sismeiro 2009) like the number of page viewed, total amount spent on the website (Bucklin & Sismeiro 2003; Johnson, Bellman & Lohse 2003) or page types (Moe 2003).

Within the field of online shopping, studies have investigated online shopping behavior to better understand and predict purchase conversion. For instance Moe and Fader (2004a, 2004b), developed stochastic models that model conversion behavior in order to predict the purchase probability of a customer. In their study, Sismeiro and Bucklin (2004) broke down the purchase process into a series of tasks and predicted the probability of completing each task and in the end purchase conversion. Similar to this study, is the study of Montgomery, Li, Srinivasan and Liechty (2004) wherein they modeled the customers’ path and found that this path is able to predict purchase conversion. However, in this study, clickstream data will not be used to predict purchase conversion, but to predict customer switch.

(15)

notifications and type of page visited. In the following sections these are discussed in more detail.

2.2.1 Total amount of website visits

The purchase process of every customer is based on four stages, namely Attention, Interest, Desire and Action (AIDA) (Strong 1925). In a competitive landscape, competition can capture the interest of the customer by the use of marketing techniques to acquire this customers for themselves. Once this attention has been captured, the customer can enter the affective stage, which include the interest and desire stage. Within this stage, the customer develops the affection to actually buy the product or service resulting in a desire. If this desire is strong enough, the customer will enter the action stage wherein the customer is taking action to buy the product or service. In this case, the customer is entering a contract with the competitive service provider. Once the customer has entered the new contract, the service provider has lost the customer. In other words, the customer has switched.

It is assumed that customers will only switch when the competitive offer outweighs the current offer. In order to decide whether this is the case, the customer has to compare both offers. This is also known as building a consideration set (Moe 2003). Building this consideration set is assumed to take place during the affective stage since competition already captured the attention of the customer but the customer is not taking actions yet to buy. One way to easily obtain the information needed to build this consideration set, is by collecting information on both service providers’ websites. This type of browsing behavior is called ‘search and deliberation’ behavior and is defined by Moe (2003) as visits that are motivated by future purchase “to acquire relevant information to help make a more optimal choice”. The information obtained is used to build the consideration set and evaluate the items in the set. In this study, the future purchase is the switch to another service provider.

The same author found in collaboration with another author, a significant relation between the customers’ amount of website visits and his or her purchase probability (Moe & Fader 2004b). Hence, the more the customer visits the website, the more likely the customer is to buy at that website. These visits were used to search and liberate the website to build the consideration set for their future purchase decision. This is similar to the expected behavior of those customer who are considering switching.

(16)

discrepancy in their findings may be explained by the difference in the type of products sold on the websites of investigation (Bucklin & Sismeiro 2009): CD’s (Moe and Fader 2004b) versus new cars (Sismeiro and Buckling 2004) . Since more financial risk is involved in purchasing a car, the customer may perform a different type browsing behavior; much website visits are the result of the need for much product information before buying the car. This type of browsing behavior is called knowledge building (Moe 2003). Hence in the case of knowledge building browsing behavior, the amount of website visits is not predictive for predicting future purchase.

Although switching between energy providers involves some financial risk in the form of a switching fine, it is still minimal as the switching fine is (partly) paid by the energy provider switched to most of the times. Therefore it is expected that, similar to the study of Moe and Faber (2004b), an increase in website visits is a sign of building a consideration set by performing search and deliberation browsing behavior and is thus predictive for customer switch. The following hypothesis is formed:

H1a: The customer’s total amount of website visits has a positive effect on the customer’s switch probability.

2.2.2 Total time spent on website

In line with thinking of the total amount of website visits, the total amount spent on the website is expected to be higher for customers who are orientating switching to another company. As their total amount of site visits will increase, the total amount of time spend on the website is expected to increase as well. Furthermore, the information gathering process to build the consideration set on requires extensive reading and browsing. According to the study of Liu (2005), people find it hard to concentrate when reading in-depth electronic documents. For that reason, when reading a webpage, they are mostly scanning for keywords which costs less time and effort. However, as switch oriented customers will look for extensive information to base their switching decision on, these customers are more likely to read the content on the page in-depth, instead of simply scanning it. As this requires a high level of concentration, this takes time. Therefore, based on these two reasons, these customers are expected to spend more time on the website than those customers who are not orientating to switch. This leads to the following hypothesis:

(17)

2.2.3 Average path length

Montgomery et al. (2004) show in their study that analyzing the sequence of the path the customer follows through the website is a useful predictor of purchase conversion. More specific, it provides information about the customer’s goal(s) of the browsing session. Which can be to buy or only to explore. This brings us back to the four types of browsing behavior found by Moe (2003), namely knowledge building, hedonic browsing, directed buying and search/deliberation. Those customer that are searching and deliberating on the website, have the highest average amount of pages viewed during one session compared to the other browsing behaviors (Moe 2003). As discussed before, those customer considering switching are expected to show search and deliberation behavior in their browsing sessions. Thus, it is expected that the average path length of this customers is higher than of those customers who are not considering switching.

Next to this, a long path can be an indication of bad website quality caused by incomplete content of the pages or bad website design (Chang & Chen 2008). This can restrain the customers’ goal of the visit. Bucklin and Sismeiro (2003) showed in their study that, similar to the offline world, customers show time constrained behavior in e-commerce. They “trade off the number of pages requested and time spent at each page” (Buckling & Sismeiro 2003). Thus the design of the webpage must be convenient so the customer can find what he or she is looking for in a minimum amount of time. In case convenience is low, the customer spends much time to find what he or she came for and at a certain point chose to exit due to time constraint. According to Chang and Chen (2008), a website with a high convenience level is one that is easy to navigate through and is user friendly. This also includes that the path must be simple and short and is found to be an important determinant of e-loyalty (Arya & Srivastava 2015; Chang & Chen 2008; Jaiswal, Niraj & Venugopal 2010). Therefore, a long path may be an indication of low website convenience and thus decreasing customer loyalty, which in turn may increase the switching probability.

Based on these arguments, the following hypothesis is given:

(18)

Website quality

2.2.4 Error notifications

An error notification is a notification a website user receives when an error occurs. The two main causes of errors occurring in web browsers are (1) crashing web browser applications or (2) user requests that do not match any web servers (Lazar & Huang 2003) Although error notifications have only been used as a covariate in the clickstream data literature (Buckling & Sismeiro 2003), in this study it is used as a predictive variable because it is believed to influence loyalty.

According to the study of Ceaparu, Lazar, Robinson and Shneiderman (2004), receiving error messages while browsing websites is one the most frustrating experiences for website visitors. Frustration is a negative emotion the customer experiences. Negative emotions are found to have a negative influence on the perceived overall service quality (White 2006). Moreover, an error message is a service failure of the website and decreases the quality and convenience of the website. E-service quality is found to be an important determinant of customer satisfaction and loyalty for both commerce and content sites by many scholars (Jaiswal et al. 2010). Therefore it is expected that it causes a decrease in the perceived service quality of the service provider resulting in a higher probability to switch.

H2: The total amount of error notifications received by the customer has a positive effect on the customer’s switch probability.

Type of page visited

(19)

product pages and additional product pages. Next to this, a more specific distinction within the service pages is made as well, namely support and self-service. All those pages that do not belong to one of these specific groups are grouped under ‘other’. The type of pages and their expected effects on customer’s switching probability are discussed more deeply below. 2.2.5 Firm Generated Content (FGC) pages

These type of pages are the pages that contain information about the company but can also contain company news updates or blogs, this type of content is also known as firm-generated content (FGC). The process that is responsible for this content is called digital content marketing. In her paper, Rowley (2008) defines digital content marketing as “the management process responsible for identifying, anticipating, and satisfying customer requirements profitably in the context of digital content, or bit-based objects distributed through electronic channels”. The aim of these pages is to create familiarity and trust among its customers (Porter, Devaraj & Sun 2013). By doing this, a deeper relationship with the customer can be developed in the way that the customer is more committed and in the end more loyal to the company (Verhoef 2003; Veloutsou 2015). More specifically, Hartmann and Ibáñez (2007), found that brand trust and familiarity are important determinants for customer loyalty in the energy market. Customer loyalty is in turn positively related to customer profitability (Hallowel 1996). Additionally, FGC is found to be positively related to customer’s spending behavior (Kumar, Bezawada, Rishika, Janakiraman, & Kannan 2016). The results of their study show that customers who were exposed to GVC through social media engagement did spend more money, showed more cross-buying behavior and were more profitable than those customer who were not exposed. Based on this, it is expected that visiting this type of pages will decrease the customer’s switching probability.

H3a: The total amount of FCG page visits by the customer has a negative effect on the customer’s switch probability.

2.2.6 Product pages

(20)

2.2.6.a Current product pages

Current product pages include the product pages about the product and/or services the customer is already owning at this moment. As discussed in section 2.2.1 ‘Total amount of website visits’, those customers that are in the purchase process (AIDA) and are thus actively thinking about switching to a competitor are expected to critically evaluate their current services and products in order to build their consideration set whether to switch or not. The information needed can be found on these current product pages and can be used for building knowledge in order to compare the offer of the competitor(s) and the current company. Therefore it is expected that visiting this type of pages is positively influencing a customer’s switching probability.

H3b: The total amount of current product page visits by the customer has a positive effect on the customer’s switch probability.

2.2.6.b Additional product pages

These type of product pages are those pages that contain information about products and services the company offers in addition to their core product or service. The aim of these type of additional product pages is to realize cross-buying in their existing customer base, which is the “the degree to which customers purchase products or services from a set of related or unrelated categories of the company” (Reinartz & Kumar 2003). Cross-buying can increase customer retention because it cross-buying causes an increase in switching costs which in turn creates a lock-in effect (Bowman & Narayandas 2001; Reinartz & Kumar 2003). By this, customer’s probability to switch is reduced and customer’s lifetime value increased (Venkatesan & Kumar 2004). Visiting these additional product pages can be an indication that a customer is in the ‘attention’, ‘interest’ or even ‘decision’ stage of the purchase process of additional products of the current company and are thus more likely to cross-buy. It is assumed that those customers who consider switching show no interest in cross-buying. They may use it as part of their consideration set but this will not be the main part and will thus not result in a high amount of visits on these pages. Contrary, those customer who are interested to cross buy will show this by a high amount of this page type visits. Hence it is expected that the more a customer visits this type of pages, the more likely this customer is to cross-buy and the less likely the customers is to switch.

(21)

2.2.7 Service pages

The service pages are those pages that aim to provide service to the customer, without direct involvement of a service employee (Kim, Kim & Lennon 2006). The following distinction in service pages is made: service support pages and self-service pages. The goals between these type of pages differ, the service support pages are there to offer customer support for questions or complaints the customer may have, while self-service pages offer an opportunity to organize the services and products the customer purchases from the company.

2.2.7.a Service support pages

The service support pages are the pages that contain supporting information about practical issues. The information is mostly provided in a question-answer manner, mostly referred to as Frequently Asked Questions (FAQ). These pages can be used for customers to find an answer to questions they have. A high visit frequency on these type of pages may indicate a high need for service support either because the customer cannot find the information the customer was looking for on the website or the planned action the customer was trying take was not possible. So the convenience of the website was low, resulting in lower customer loyalty (Arya & Srivastava 2015; Chang & Chen 2008). Moreover, service support pages mostly contain information about issues that must be handled by the company like complaints or ending service contracts. It is assumed that customers who have complaints and thus are dissatisfied will visit these type of pages more than those customers that do not have complaints. As dissatisfaction is negatively influencing customer loyalty it is expected that customer that visit these type of pages often are more likely to switch than those customer that do not visit these that often.

H3d: The total amount of service support page visits by the customer has a positive effect on the customer’s switch probability.

2.2.7.b Self-service pages

(22)

more likely to switch and thus these visits have a positive effect on the relationship between online behavior and customer switch.

H3e: The total amount of self-service page visits by the customer has a positive effect on the customer’s switch probability.

2.2.8 Other pages

These are the pages that do not belong to one of the page types described above, or a combination of the page types. As the content of these pages are highly differing and can therefore not be interpreted in their effect on customer switch, no relationship between these pages and customer switch is expected.

2.3 Customer service call behavior

Service call center

Until today most service call center literature has focused on the organizational side of the service, like increasing employee productivity (Singh 2000) or personality factors increasing employee performance (Sawyerr, Srinivas & Wang 2009). Some other research focused on the service call agent behavior and how this behavior is perceived by the customer (Burgers, Ruyter, Keen & Streukens 2000; Ruyter & Wetzels 2000). Burgers et al. found that the customer expects the service call agent to be adaptive, create a high level of assurance, show empathy and authority. When these expectations are met, this creates a feeling of satisfaction among the customer. These finding inspired other scholars to investigate this effect further and linked in the service employee behavior to customer loyalty. The findings of these scholars formed the basis for the expected relationship between service call behavior and customer switch and is discussed in the next section 2.3.1 ‘Incoming service call’.

2.3.1 Incoming service call

(23)

and responsiveness. They found that attentiveness is directly resulting in satisfaction while perceptiveness was found to create trust. The last type, responsiveness, was found to influence both positively. A similar study by Dean (2009), found that customer focus of the service call center is directly linked to customer self-reported loyalty. Within the service industry, Levy (2014) found that for those customers who have a high level of service usage personal interaction with the company increases their level of commitment towards the company. Thus is seems that the personal nature of the service call center can increase customer loyalty. Therefore it is expected that those customers who have had contact with a service employee by phone during the observation period, are less likely to switch.

H4: Customer’s contact with the service call center has a negative effect on the customer’s switch probability.

2.4 Control variables: customer characteristics and relational characteristics

Within many studies customer characteristics and relational characteristics are found to be important antecedents for customer behavior (Prins & Verhoef 2007) and customer retention (Verhoef et al. 2007) and should therefore be controlled for. Based on previous literature this study controls for the following customer characteristic: age and the following relational characteristics: usage of service, additional products, additional charges, length of relationship. These are discussed more comprehensively below.

2.4.1 Age

Within several studies, age is found to be an important determinant of customer switch (Coussement & Van Den Poel 2008; Homburg & Giering 2001; Wong 2011). The higher the age in years, the less likely the customer is to switch. Therefore the age of the customer is controlled for in this study.

2.4.2 Type of contract

(24)

2.4.3 Usage of service

Within the energy sector, usage of the service is found to be an important determinant of customer switch (Wieringa & Verhoef 2007). Those customers that use much energy resulting in high monthly costs, are more critical and more likely to compare energy providers or even switch provider. Therefore the usage of the service is controlled for.

2.4.4 Additional product

Within the energy service industry additional, value-adding products can be bought. These can be bought in addition to their energy contract which is the core service. With the similar reasoning as type of contract, it is expected that those customer who have bought the additional product are less likely to switch.

2.4.5 Additional charges

Within the service industry contracts with the customer are used to manage the relationship with the customer. The contract determines how much of the service is provided to the customer, for instance how much energy or telephone minutes. When the customer uses more than is stated in the contract, the customer has to pay an extra fee. This can cause dissatisfaction at the customer and can trigger switch behavior. Therefore, additional charges are controlled for.

2.4.6 Length of relationship

(25)

Figure 1: Conceptual Model

Chapter 3: Methodology

3.1 Data collection

3.1.1 Sunwinergy

As this study is interested in within-website behavior, site-centric data has been used. The data is collected from the website of Sunwinergy, a Dutch energy provider. This energy provider provides energy to approximately 4.4 million customers whereof 1.5 million individual customers. The other 3 million customers are business customers. In this study the B2B contracts are left out of scope. The website of Sunwinergy consists of the page types described in 2.2.5 ‘Type of page visited’.

3.1.2 Original Dataset

(26)

2015. The server does only allow to log visitors data of those that are logged into their online account, called ‘Mijn Sunwinergy’. Therefore it was only possible to analyze those customers who were logged into their ‘Mijn Sunwinergy’ account. Hence, the online behavior of those customers who were not logged into their online account during the observation period is unknown. The consequence of this is that a self-selection bias was present. This bias is further discussed in chapter 6 ‘Limitations and Further Research’. The original clickstream dataset contained data on a session level meaning that for all the customer’s sessions the visited URL’s and their belonging data- and timestamps were available. After cleaning the data (discussed in 3.2 ‘Data Cleaning’), this dataset was aggregated to a higher level wherein the website behavior was aggregated to a customer level meaning that general online behavior (the total amount of website visits, the total time spent on the website and the average path length) remained. After this aggregation process, switching data, service call center data and customer characteristics data were merged into the one dataset. This resulted into a dataset consisting of a total amount of 131,551 visiting customers.

The switching data was measured over a period of 120 days from 1st_{of September 2015 until} the 31st of December 2015. This period was chosen for two reasons: firstly this period provides a realistic range for marketers to develop retention campaigns for those customers that have a high probability to switch. Secondly, Sunwinergy’s current switching model is based on this period as well and can therefore be easily extended with the predicting results found in this study. A customer is registered as switched once the competitor has informed Sunwinergy about the switch. Other churn behavior like movement or death was ignored as this study is particularly interested in customer switch.

For the service call center data was chosen to only measure whether or not the customer has had contact with the service call center. The reason for this is that the call agents log the data about the subject and content of the call themselves and was found to be rather subjective which would could cause unreliable results. The service call data was measured over the period from the 1st of March until the 31st of August.

3.2 Data Cleaning

3.2.1 URL selection

(27)

continue studying the top 200 URL’s of this ranking. Some of these URL’s included administrative pages that are highly correlated with the page types like self-service and product pages. Because after executing a certain action on one of these pages, a confirmation URL is given. For example, when changing a password, the server leads the customer automatically to a new generated page where a confirmation message stating that the password is changed is provided. These are not the type of pages of interest in this study and do not tell much about the underlying behavior of the customer. Moreover, the aim of this study is to investigate Sunwinergy’s individual customers (households) behavior and not their business customers. Although only those customers with a B2C contract were selected, a few customers had visited the business area of the Sunwinergy website. As this part of the website is not of interest in this study, these URL’s were deleted from the dataset as well. After this cleaning a total amount 169 of unique interpretable URL’s remained. An overview of these URL’s is given in Appendix A.

3.2.2 Outliers, oddities and missing values

After selecting the URL’s, the dataset was investigated for outliers, oddities and missing values.

One of the issues found was the issue that some customer visits were not correctly logged by the server resulting in either unknown visiting times or the same URL logged in sequential timestamps. The first issue of unknown visiting times resulted into the problem that it was not possible to calculate the total time spent during this session. For these sessions the missing value on total time spent on the website was filled by the outcome of the following multiplication: (the average time spent per page * (the amount of pages – 1)). The minus one is included to compensate for the unknown time spent on the last page (further discussed in chapter 6 ‘Limitations and Further Research’). The second issue caused the path length of that particular session to be extremely long and by this increasing the customer’s average path length. For these customers their overall average path length was replaced by the overall average path length of all customers.

(28)

Moreover, several outliers were detected. Firstly, some customers had visiting the website more than 200 times according to the data resulting in an extreme high score on total website visits. Interpreting these scores indicate that the customer would have visited the website around one time per day during the observation period. These high amounts of sessions are not very realistic and are assumed to be caused by Sunwinergy employers who use the website on a daily basis for their jobs and are logged into their private ‘Mijn Sunwinergy’ account. For this reason, these customers were deleted form the dataset. The remaining outliers that had a score above 100 total visits were replaced by the average scores plus 3 times the standard deviation (mean + 3* standard deviation), in order to normalize the deviation of the variable. Furthermore, some extremely high outliers were found in the amount of minutes spent on the website. Those customers with a total time spent on the website higher than 1000 minutes were found to be the same customer as those with the total amount of website visits higher than 200 visits and thus were deleted from the dataset.

For the variables ‘average path length’ and ‘usage of service’ some outliers were detected as well. For both variables the outliers were replaced by their mean plus three times the standard deviation (mean +3* standard deviation), again with the aim to normalize both variables.

A few missing values appeared in the dataset for the variables ‘service usage’ and ‘incoming service call’. For those customers with an unknown usage of service level, these missing values were filled by the average yearly usage of customers with the same contract (electricity, gas or both). Those missing values for ‘incoming service call’ were replaced with a 0, indicating no contact. All calls to the service call center are logged and therefore it can be assumed that these customers simply did not have contact with the service call center during the observation period.

3.3 Sample set

3.3.1 Stratified random balanced sample set (switch vs non-switch)

(29)

no longer a rare event and increases the predictive accuracy of the model (Blattberg et al. 2008). As this study made use of a logit model, this procedure did not influence the fit of the parameters estimates except the constant (Donkers, Franses & Verhoef 2003; Scott & Wild 1997). As the value of the constant is not of interest in this study, no further actions were taken to correct for this. Another issue resulting from the selective sampling is that the true standard errors will be smaller than the ones based on the full sample (Donkers, Franses & Verhoef 2003). For the same reason as the constant, no further corrections were applied. An overview of the original sample and the balanced sample size is given in Table 1 ‘Balanced Sample’.

Table 1: Balanced Sample

Dependent variable Description Amount in original sample size Percentage of original sample size Balanced sample size Percentage of balanced sample size Switch

Those customer that did switch to another energy provider within the period September 2015 until December 2015 (120 days). N=3994 3.04% N=3994 50% Non-switch

Those customer that did not switch to another energy provider within the period September 2015 until December 2015 (120 days).

N=127557 96.96% N=3994 50%

(30)

biggest amount of switchers result from those customers with a combined contract. This is because around 85.9% of the original dataset consisted of customers with combined contracts, around 11.4% electricity-only contracts and 2.7% gas-only contracts. But when looking at the percentage switch relative to the customer with the same contract type, customers electricity-only are significantly less switching than the overall switching rate of 3%. The combined contracts and gas-only contract are having switching rates almost similar to the overall switching rate. Based on this information it was chosen to ensure that the same ratio of the type of contracts was represented in the balanced sample, instead of a randomly chosen group of non-switching customers. Thus both within the switch and non-switch group, the combined contract group was represented in an amount of 3,432 (85.9%) customers, electricity-only contract with an amount of 456 (11.4%) customers and gas-only contract with 106 (2.7%) customers, making a total of ,3994. By this it was prevented that one of the groups would be omitted by the random sampling. The balanced sample defined before has now become a stratified random balanced sample set. Herein the strata are the contract types but the customers within these strata are still randomly chosen. The sample is balanced as the amount of switchers and non-switchers is balanced equally.

Table 2: Switching per Contract Type Contract type Amount of customers Switching customers within category % of total switch % switch relative to contract type Combined (E&G) 11,3013 3,554 89% 3.1% Electricity-only 15,044 306 7.70% 2% Gas-only 3,494 134 3.40% 3.4% Total 131,551 3,994 100% 3% 3.4 Measurement approach

In the following part a description of the measurement approach of each variable is given.

3.4.1 Total amount of total website visits

(31)

study of Bucklin and Sismeiro (2003), a session was counted as a website visit when more than one page was visited during the session. Those visits including one page visit only are not considered as a browsing session and does tell much about the online behavior of the customer.

3.4.2 Total time spent on website visits

The total amount spent on the website was measured by the sum of the time spent per customer web browsing session on the Sunwinergy website. The time spent per session was calculated by calculating the difference between the entering time registered of the last page and the first page. As Sunwinergy’s webserver log files could not record the exit time of the visitor, the exit time of the last page was unknown and therefore the total time spent on this page could not be measured. The consequences of this limitation is further discussed in chapter 6 ‘Limitation and Further Research’.

3.4.3 Path length

The customer path length was calculated by taking the average of path length over all the customer’s sessions. The minimum amount of pages for the path length was two as only those visits with more than one page visit were included in the study (discussed in 3.4.1 ‘Total amount of total website visits’).

3.4.4 Total amount of error notifications

An error notification appeared in the dataset as an URL containing the error codes 404,500 and ‘error notification’. The total amount of these error codes per customer per browsing session was counted and was calculated as a percentage of the total amount of pages visited.

3.4.5 Type of page visited

(32)

3.4.6 Incoming service call

The incoming call was measured as whether the customer has had contact with the customer service call center. It was measured as a dummy with 0= no incoming call and 1 = incoming call.

3.4.7 Control variables: customer characteristics

The control variables in this study were customer characteristics. The first measured customer characteristic was the age of the customer. The age of the customer was measured in years in a range from 18 to 95 years old. This was the age of the customer registered at Sunwinergy. So the age of the rest of the household was unknown. Within the dataset 3 type of customers existed: first customers that bought both electricity and gas (combined contract), secondly customers that only bought electricity and lastly customers that only bought gas. These contract types were measured by the following categories respectively: 0 = combined contract (electricity and gas), 1 = electricity only and 2 = gas only. The usage of the service was measured as an index which compared the customer’s yearly usage of the energy service with the average usage of other customers with the same contract. Thus, those customer with an electricity contract had an index wherein the yearly energy usage in kWh was compared to the average yearly usage (=3134.324 KwH) of electricity of all customers with only an electricity contract. Similar, this index was also calculated for gas usage (in m3). Herein the average yearly gas usage was 1,577.835 m3 per customer. For those customers that had a combined contract (both electricity and gas), the sum of the customer’s yearly electricity and gas usage was calculated and divided by the sum of the average of yearly electricity and gas usage of all customers with a combined contract (=4463.311). During the measured period it was only possible for customer who had an energy contract at Sunwinergy to buy the additional product, which is a smart temperature device called TOON. This is a home energy device with allows customer to gain insights into their energy usage of costs. Whether a customer had bought this device from Sunwinergy was indicated by a dummy 0=no, 1=yes. Additional charges were measured as whether or not the customer had to pay additional on his/her year note. This was indicated by a dummy with 0=no additional payment, 1= additional payment. The last control variable ‘length of relationship’ was measured in the amount of months a customer was in a contractual relationship with Sunwinergy.

3.4.8 Customer switch

(33)

An overview of the variables measurement approach is given in Table 4 ‘Variables Measurements Approach’ below.

Table 4: Variables Measurement Approach

Independent variables Description Notation

Total amount of website visits

The customer’s total amount of website visits in URL’s.

Numeric / scale

Total time spent on website visits

The customer’s total time spent on the website in minutes.

Numeric / scale

Average path length The customer’s average path length per session.

Numeric / scale

Total amount of error notifications

The total amount of error notifications the customer has received measured in percentage of the total amount of website visits.