WHICH CAME FIRST, THE CHICKEN OR THE EGG?

WHICH CAME FIRST, THE CHICKEN

OR THE EGG?

INVESTIGATING THE CAUSAL RELATIONSHIP BETWEEN LOYALTY AND

VALUE-ADDING SERVICES IN A SERVICES MARKET

A. R. Mevissen

University of Groningen

WHICH CAME FIRST, THE CHICKEN

OR THE EGG?

INVESTIGATING THE CAUSAL RELATIONSHIP BETWEEN LOYALTY AND

VALUE-ADDING SERVICES IN A SERVICES MARKET

Master Thesis Marketing Intelligence & Marketing Management January 2016

Ate Ruben Mevissen, s1892282 Waterloostraat 142A01, 3062 TW Rotterdam

atemevissen@hotmail.com +31 (0)633099281

1s t Supervisor: dr. J. (Jelle) T. Bouma (j.t.bouma@rug.nl)

2n d Supervisor: dr. ir. M. (Maarten) J. Gijsenberg (m.j.gijsenberg@rug.nl) External supervisor Eneco: M. (Marijn) Pannekoek (marijn.pannekoek@eneco.com)

University of Groningen Faculty of Econo mic & Business

EXECUTIVE SUMMARY

Over the past two decades, the energy market in The Netherlands has changed significantly. Key changes include the liberalisation of the market, the decline in switching costs, changing consumer demands and technological changes. These changes have radically transformed the view on loyalty and simultaneously increased the importance of the concept. The recent adoption of more

sustainable and relationship-oriented business models by large companies in the energy market adhere to this appeal for more consideration towards loyalty. As a result, a series of value-adding services and products has been introduced to the market. Due to the recent introduction of the services in the energy industry, little is known about the effect of these services on loyalty and, conversely, the effect of loyalty on these services in this specific context. Therefore, the following research question was developed ‘What is the causal relationship between the adoption of

value-adding services and behavioural loyalty and what is the mediating effect of attitudinal loyalty in this causal relationship?’. The importance of answering this question is to gain familiarity with the

phenomena and acquire new insights, from a theoretical services marketing perspective. From a managerial perspective, answering this question is important to find establish where and how marketing budgets should be spent.

Loyalty is separated into two constructs, behavioural loyalty and attitudinal loyalty. As a result, the relationships between behavioural loyalty and value-adding services are analysed directly and attitudinal is tested as mediating variable within these relationships.

For this explorative research, two models were developed comprehending both relationships and for both models a multinomial logit regression analysis was performed. Overall, this study reveals that the relationship between behavioural loyalty and value-adding services is indeed bi-directional of nature but no dominant direction can be assigned. Moreover, it is concluded that attitudinal loyalty does not play a mediating role in this relationship, in either direction. In addition, several managerial implications result from this study. First of all, companies should include one or more services in their product portfolio as the adoption of such services increases the lifetime value of customers. Likewise, the results suggest it is recommendable to up-sell customers to services which add more value as this results in a higher customer lifetime value. Additionally, the absence of mediating power of

TABLE OF CONTENTS

2.1.3 Loyalty in a services market ... 9

2.2 Value-adding services ... 10

2.2 Loyalty and value-adding services ... 12

2.2.1 Influence of value-adding services on loyalty ... 12

2.2.2 Influence of loyalty on value-adding services ... 13

2.3 Control variables ... 15

2.4 Conceptual research model ... 16

3. DESIGN ... 17 3.1 Data collection ... 17 3.2 Measurement approach ... 17 3.2.1 Value-adding services ... 17 3.2.2 Behavioural loyalty ... 18 3.2.3 Attitudinal loyalty ... 18 3.2.4 Control variables ... 18 3.3 Model specification ... 18 3.3.1 Model 1 ... 19 3.3.2 Model 2 ... 20 4. RESULTS ... 22 4.1 Model 1 (2013) ... 23 4.1.1 Base analysis ... 23 4.1.2 Mediation analysis ... 29 4.2 Model 2 (2015) ... 31 4.2.1 Base analysis ... 31 4.2.2 Mediation analysis ... 34 4.2.3 Hypothesis table ... 35 5. CONCLUSIONS ... 37 5.1 Theoretical implications ... 37

5.1.1 Influence of value-adding services on loyalty ... 37

5.1.2 Influence of loyalty on value-adding services ... 39

5.1.3 Causal relationship ... 41

5.2 Managerial implications ... 42

5.3 Limitation and directions for future research ... 43

6. REFERENCES ... 44

7. APPENDICES ... 60

APPENDIX 1: Correlations statistics ... 60

1

1. INTRODUCTION

Over the past two decades, the energy market in The Netherlands has changed significantly. This change started with the liberalisation of the market in 1998, initiated by the Dutch government which acted upon two EU electricity directives that established common rules for generation, transmission and distribution of electricity. Before the liberalisation, the market consisted of 23 licensed distribution companies, owned by municipalities and provinces, which were obliged to supply electricity in their respective geographical areas where they acted as monopolists (van Damme, 2005). After the Dutch government passed the electricity law in 1998, many of these

companies privatised and merged, resulting in three large players: Nuon, Essent and Eneco (Wieringa & Verhoef, 2007). In 2004, the energy market, including the green energy market, was completely liberalised, allowing consumers to freely choose their supplier of electricity and gas.

The liberty to freely choose a supplier led to another large change within the energy market. Consumer switching costs declined, thereby lowering the entry barriers to the market resulting in opportunities for new market entrants (Porter, 1979). Before the liberalisation of the market, the companies acting under the name of the three incumbent suppliers served 90% of the households in The Netherlands (Pomp & Shestalova, 2007). Since then, numerous new entrants seized a portion of this substantial market share of the incumbent companies. Mainly discounters and ‘green’ energy providers offering low prices and sustainable energy entered the market. In less than a year after the opening of the market, two of the largest market entrants, Oxxio and Greenchoice already accounted for 8% of the total market share (DTe, 2006). As a result of the new entrants, 11 years after the liberalisation of the energy market the three incumbent suppliers (without acquired competitors such as Energiedirect.nl, Oxxio and Woonenergie) have seen their combined market share decline to less than 70% (GfK Energiemeter, 2014). Converging market shares of smaller and larger suppliers are likely to continue as the percentage of consumers who switched to another energy and gas supplier has been increasing over the past decade from 5% in 2006 to 13% in 2014 (ACM, 2015). In addition, since the liberalisation of the energy market almost half of the consumers switched at least once (ACM, 2015). The high degree of switching behaviour suggests a low level of loyalty towards the energy suppliers as switching behaviour is associated with monetary, time, effort and relational investments motivating customers to stay in an existing relationship with a company (Dwyer et al., 1987; Heide & Weiss, 1995; Burnham, Frels & Mahajan, 2003).

2

that 66% of the people around the globe are willing to pay more for sustainable brands. In addition, consumer energy consumption is under pressure in many Western countries such as The Netherlands (Energie-Nederland, 2014). As a result, the need for energy minimisation increased. This need for energy minimisation together with technological advances paved the way for home energy products and service such as the smart meter which form a platform for combining energy usage and feedback (Froehlich et al., 2011). Hence, consumers will be able to gain insights on their home energy usage. These insights will aid to minimise energy consumption and allow consumers to contribute to a more sustainable world. More notably, it reduces the consumers’ energy bills which is increasingly

important as price has been and remains to be the number one reason to be disloyal towards an energy supplier (Meijer & Perfors 2004a, 2004b; GfK energiemeter, 2014; ACM, 2014).

Even though price remains the number one incentive to stay at or leave an energy provider on the consumer side, on the supplier side the margins for energy and gas are steadily decreasing partly due to decreasing demand and overcapacity. As a result, large actors in the energy sector including Essent, Eneco and the German energy producer RWE increasingly divert their effort towards another, more sustainable business model where focus is diverted largely from competition on price towards competing on service provision through a diversified product portfolio (e.g., fd.nl, 2014). Besides electricity and gas, energy producers and suppliers will focus on renewable energy sources such as wind and solar energy and, on product level, products and services such as boilers are offered more prominently to increase cross-selling.

The changes faced by the Dutch energy market described above have impacted the way the concept of loyalty is approached in this context. Prior to the liberalisation, loyalty was irrelevant as consumers received their energy from specific providers as this was geographically fixed. After the liberalisation, this market model completely burst. Energy providers suddenly had to compete to acquire new customers and retain current ones and the role of loyalty grew more prominently. For a long period of time, loyalty in the energy market could be confined to behavioural loyalty as the focus was largely directed at acquiring and retaining customers for as long as possible. The extensive switching

3

loyalty. Attitudinal loyalty proves to be of increasing importance as low energy bills are no longer the only motivation to remain at a current energy provider. This is in line with Dick & Basu (1994) who concluded that behavioural loyalty is paramount when describing loyalty, but without attitudinal loyalty it is incomplete. Hence, the major changes faced by the Dutch energy market have radically changed the view on loyalty and simultaneously increased the importance of the concept.

The recent adoption of more sustainable and relationship-oriented business models by large companies in the energy market adhere to this appeal for more consideration towards loyalty. As a result, a series of value-adding services and products has been introduced to the market. In this context, these products and services will provide customers with insights to manage their personal energy usage. Such products and services increase the engagement of customers as they raise awareness through interaction with the company and consumer energy usage (Darby, 2010;

Chakravarty & Gupta, 2013). By doing so, a more sustainable and lasting customer base is generated. This allocation of expenditures is in line with Reinartz et al. (2005) who reason that retention

expenditures have a greater impact on long-term profitability than acquisition expenditures when applied correctly. Also, Gupta et al. (2004) acknowledge this by finding evidence that a 1%

improvement in retention can result in a 5% customer profitability compared to 0.1% for similar improvements in acquisition. As a result, behavioural loyalty is increased through the use of attitudinal loyalty of the current customer base rather than spending large sums of money on acquiring new customers by increasing customer engagement and satisfaction. By offering these services and products, companies aim to build a stronger relationship with their customers and retain a large customer base for longer periods of time by stopping the yearly or biyearly switching behaviour amongst customers in line with previous studies (Hart, Heskett & Sasser, 1990; Reichheld & Sasser, 1990).

4

their brand and bolster this company-customer relationship into a more desired relationship, which could be, for example, long-term oriented rather than short-term oriented (Avery et al., 2014). In addition to increased behavioural loyalty, attitudinal loyalty is increased as more positive attitudes towards that company are developed through this improved loyalty. Reversely, more loyal customers are known to be more inclined to buy additional products from the same company (Reinartz et al., 2008). If positive attitudes and commitment towards a company grow, customers are more inclined to purchase or adopt additional products and services from this same company. Therefore, it could be that not the adoption of value-adding services increase loyalty but loyalty increases the adoption of these specific services. This leads us to the ancient ‘chicken and egg’ causality dilemma: which concept holds what role in this cause-effect relationship? Based on the discussion above, the following research question was developed:

‘What is the causal relationship between the adoption of value-adding services and behavioural loyalty and what is the mediating effect of attitudinal loyalty in this causal relationship?’

Several sub questions will assist in answering the research question above by focusing on specific details. These are the following questions:

1. What is the effect of the adoption of value-adding services on behavioural loyalty?

2. What is the role of attitudinal loyalty in the effect of the adoption of value-adding services on behavioural loyalty?

3. What is the effect of behavioural loyalty on the adoption of value-adding services? 4. What is the role of attitudinal loyalty in the effect of behavioural loyalty on value-adding

services?

The focus of this paper will be in the context of the Dutch energy market. It will focus on Eneco, a major player in the Dutch energy market. Customer data collected by Eneco over the past years will be used for analysis. Moreover, competitive variables will not be included in this research.

5

companies to better assess how consumer beliefs about brand attributes influence their brand purchase behaviour. Secondly, this paper will provide more clarity in more efficiently allocating marketing budgets. By tackling different aspects of the cause-effect issue, clearer starting points can be adopted. Hence, should marketing budgets be spent on promotional efforts advocating the adoption of value-adding services or should they be spent on a specific customer segment with a certain degree of loyalty. Also, companies will be able to assign a relative value to each of the services categorised by degree of additional value they provide. These insights will provide companies with guidelines on how much effort they should put in to acquire or retain a customer. Lastly, companies will be able to design a completer customer journey by incorporating better fitting, more personalised customer touchpoints through which the customer-company relationship can be strengthened. The relevance for literature is threefold. First of all, the causal relationship between value-adding services and loyalty within a commodity services market will be established. The outcomes will provide researchers with initial insights and a starting point for future research. Secondly, diverting the focus from the technical properties of value-adding services in the energy market to insights for consumer behaviour from a service marketing perspective. To date, a great deal of attention has been directed towards the technical properties of these services, rather than the implications for services marketing literature (Darby, 2010). As a result, this study will contribute to the current literature by providing insights of these services in relation to consumer loyalty. Moreover, Bandyopadhyay & Martell (2007) have found that attitudinal loyalty affects behavioural loyalty and is an antecedent of behavioural loyalty. The researchers studied the toothpaste market and ask for support from other products, service and markets to validate the bi-dimensionality of customer loyalty. This paper will provide insights on the (non) bi-dimensionality of customer loyalty from a services marketing perspective.

6

2. THEORETICAL FRAMEWORK

This paper studies the causal link between loyalty and value-adding services. First, this chapter will elaborate on the relevant facets of loyalty including behavioural and attitudinal loyalty before the link between them. Next, value-adding services will be elaborated upon and linked to loyalty. Lastly, several variables are elaborated upon which are to be controlled for. Throughout this chapter, multiple hypotheses are proposed and to conclude with a conceptual model is developed.

2.1 Customer loyalty

Customer loyalty towards a brand is defined by Oliver (1999) as:

‘A deeply held commitment to rebuy or repatronise a preferred product/service consistently in the future, thereby causing repetitive same-brand or same brand-set purchasing, despite situational influences and marketing efforts having the potential to cause switching behaviour’.

This definition provides a distinction between two aspects of loyalty: behavioural and attitudinal loyalty. There has been a plethora of studies in the field of brand loyalty over the past decades which failed to provide a conclusive definition of consumer loyalty which means the above definition may not be accepted by everyone. Many studies have taken a unidirectional behavioural perspective, taking repeat purchasing behaviour as the focal construct in defining consumer loyalty (e.g., Cunningham, 1966; Massey et al., 1970; Kahn et al., 1986; Ehrenberg et al., 1990; Neal, 1999). Another, more recent, line of thought adopted in many other studies (e.g., Day, 1969; Lutz & Winn, 1974; Jacoby & Chestnut, 1978; Baldinger & Rubinson, 1996; Thiele & Bennett, 2001; Rundle-Thiele, 2005) questions this behavioural approach. These studies argue that repeat purchasing behaviour is too limited and is incomplete in describing customer loyalty. They add an attitudinal dimension to the consumer loyalty construct to make it a more holistic, bi-directional approach. Below, both interpretations of loyalty will be discussed and, finally, they will be linked.

2.1.1 Behavioural loyalty

As mentioned above, many studies have taken a behavioural approach in defining loyalty (e.g., Cunningham, 1966; Massey et al., 1970; Kahn et al., 1986; Ehrenberg et al., 1990; Neal, 1999). Nonetheless, defining behavioural loyalty is difficult as it varies in different contexts. This difference is largely attributable to market characteristics such brand switching, purchase frequency, share of category, perceived risk and inertia (Rundle-Thiele & Bennett, 2001).For example, in a consumables market such as the cereal or ice cream market, a logical definition of loyalty is a ‘buyer of only one

7

happens often, sometimes even within a single shopping trip. On the other hand, in service markets behavioural loyalty can be defined as ‘purchase behaviour repeated over time’ (García Gómez et al., 2006). This is a appropriate definition as multi-brand buying often is not an option and buying frequency is low. The definition by Garcia Gómez et al. (2006) is the definition which is adopted by this paper. As with defining behavioural loyalty, developing correct measurements differs across contexts as well. Different approaches towards behavioural loyalty have been taken including proportion of purchase (e.g., Cunningham, 1966; Neal, 1999), purchase sequence (e.g., Kahn et al., 1986) share-of-wallet (Berger et al., 2002; Mägi, 2003) and probability of purchase (Massey et al., 1970). Another, more recent, approach adopted by literature to measure behavioural loyalty is the customer lifetime value, from hereinafter CLV (Reinartz and Kumar, 2000; Rust, Zeithaml & Lemon, 2000; Kumar & Shah, 2004; Rust, Lemon & Zeithaml, 2004). CLV measures the economic worth of a customer (Berger & Nasr, 1998). More elaborately, CLV represents value of all current and future profits generated from a customer over the life of his/her business with the firm. The focus is on the individual customer and explicitly incorporates the possibility of customer defection to a competitor in the future (Gupta & Zeithaml, 2006). Also, CLV adheres to the essential idea in modern marketing that customers are assets (Rust et al., 2004). Companies must therefore invest in terms of marketing expenditures to the current customer base to ensure future revenues for the firm (Shugan, 2005). The preference over other metrics is due to the forward looking ability of the metric as it captures all the components of revenues, expenses and customer behaviour. It combines the majority of the approaches mentioned above in one single metric. Also, CLV is known to outperform and replace some of these traditional metrics (Reinartz & Kumar, 2000). A difficulty regarding CLV, however is the fact that it is difficult to measure and out of bounds for the majority of the companies to acquire partly due to missing data problems (Rundle-Thiele & Bennett, 2001; Stahl, Matzler, and

Hinterhuber, 2003; Berger et al., 2006)

As mentioned above, describing loyalty from a behavioural perspective is considered by some as rather limited and has faced significant of resistance. Repeat purchasing may, for example, be caused by a lack of preferences, lack of brand awareness or action inertia (Jacoby & Chestnut, 1978; Oliver, 1999; Macdonald & Sharp, 2000). None of these signify true behavioural loyalty but rather a

8

2.1.2 Attitudinal loyalty

Chaudhuri & Holbrook (2001) define attitudinal loyalty as ‘a degree of dispositional commitment in

terms of some unique value associated with the brand’. This definition extents the idea of

behavioural loyalty by assigning a certain value to the brand. Oliver (1997) proposed this as well by stating attitudinal loyalty develops over time through three distinct phases before the behavioural loyalty is reached. The first phase is cognitive loyalty where the customer gets familiar with the brand and preference is indicated towards the brand. Loyalty towards the brand is based on experience-based information and still rather shallow. When the task is routine and satisfaction is not processed, loyalty is not more than mere performance. If satisfaction is processed, loyalty develops towards an affective phase. This is also discussed by Geyskens et al. (1996) stating that a customer has grown into liking the relationship because it likes the partner and enjoys the partnership and desires to continue it because it experiences a slight sense of loyalty.The second phase is characterised by liking or attitude towards the brand, accumulated by multiple encounters with the brand. Commitment towards the brand is stronger at this phase and is not as easily challenged with counterarguments compared to commitment in the cognitive phase. Nonetheless, in this phase consumers still switch despite their prior satisfaction. The next phase, the conative phase, overcomes this switching as a result of brand-specific commitment to repurchase. Multiple encounters with the brand causing positive affect towards the brand grow a behavioural intention towards the brand. This deep commitment might resemble deep intentions to repurchasing, however it fails to convert these intentions to actions. The last stage redirects us back to behavioural loyalty which is the translation from intentions towards actually purchase behaviour. The framework by Oliver (1997) merges the two incomplete facets of loyalty described above into one comprehensive explanation of consumer loyalty towards a brand. Moreover, it provides us with a path from no loyalty to

purchasing behaviour. Dick & Basu (1994) add to this that social and situational factors may influence behavioural loyalty. These factors may either support or contradict an attitude and exist in the form of, for example, social norms or in-store promotions. For example, a consumer may intent to

purchasing a certain brand however, this brand may be perceived as rebellious or cheap resulting in a lack of actual purchasing behaviour or a consumer intending to purchase a certain product but decided to purchase a competitive product as result of in-store promotions. Dick & Basu (1994) argue that a stronger relative attitude will overcome these damaging social and situational factors. Several studies find evidence for the fact that attitudinal is an antecedent and has a positive effect on behavioural loyalty (Keller, 1993; Bandyopadhyay & Martell, 2007).

Developing a measurement of attitudinal loyalty as a part of consumer loyalty has been attempted frequently over the years (Berne, Mugica & Yagu, 1997; Mittal & Lasser, 1998; Chaudhuri &

9

towards a brand by measuring the perceived consumer overall rating of the brand. Various loyalty metrics exist to measure the overall brand rating. Measures that have been proposed are average customer satisfaction and repurchase intentions (e.g., Fornell, 1992; Anderson et al. 1994, 1997), the Net Promoter Score (Reichheld, 2003), intention to repurchase (e.g., Mittal et al., 1998; Kamakura et al., 2002) and ‘Top-2-Box’ satisfaction score (e.g., Myers, 1999; Morgan et al. 2005). These loyalty metrics measure the perceived consumer overall rating of a brand in relation to performance of the firm. In the past, Morgan & Rego (2006) have attempted to compare these different measures in relation to the firm performance and part of the outcome was quite surprising. For example, as opposed to Reichheld (2003) and common belief in practice, the authors came to the conclusion that the customers’ recommendation intentions and behaviours (e.g. Net Promoter metrics) had no substantial power in predicting future firm performance. Another conclusion by Morgan & Rego (2006) is more straightforward. Their results assign the highest predicting value of loyalty and, subsequently, firm performance to the average satisfaction score which is in line with other studies (e.g., Fornell, 1992; Rust et al., 1994, 1995). More recent research by De Haan et al. (2015) research disagrees with Morgan & Rego (2006) as they provides us with evidence that combining the NPS and ‘top-2-box’ satisfaction metrics provides the best predictive power of a model compared to the other metrics.

2.1.3 Loyalty in a services market

The focus of this paper is directed at the energy market. Being a services market, the energy market faces different circumstances compared to a products market which cannot be generalised

(Keaveney, 1995). Services markets are often characterised by the difficulty of evaluating the services (Javalgi & Moberg, 1997). Hence, the influence of perceived risk is greater and may act as a barrier to switching behaviour (e.g., Zeithaml, 1981). Moreover, the importance of relationships and

10

customers. On the one hand, the low degree of involvement allows the incumbent suppliers to retain a relatively stable customer base due to action inertia which is difficult to overcome for firms

(Colgate & Lang, 2001). For many consumers action inertia eventually leads to lock-in as their

propensity to search for alternatives decreases (Zauberman, 2003). On the other hand, it is extremely difficult to retain customers because of the difficulty to grow loyalty among the customer base due to the low degree of involvement. Oliver (1999) even claims some firms cannot attain the ultimate loyalty state. Requirements to achieve this are product superiority (as the minimum), determined brand defenders and a supportive social environment. Since energy suppliers mainly sell products which are homogenous across the market, product superiority is difficult to attain and therefore involvement as well. One option to overcome this would be the introduction of additional value-adding offerings which provide these services to differentiate from other suppliers and reach product superiority. Value-adding services will be discussed in the next paragraph.

2.2 Value-adding services

Before discussing value-adding services, one needs to be aware of the underlying meaning of ‘added value’ and its position in the consumer loyalty literature. Many authors have recognised the difficulty of defining value (e.g. Piercy and Morgan, 1997; Woodruff, 1997) as the definition is subjective (Hardy, 1987), varies between customers (Wikström & Normann, 1994) and within customers (Parasuraman, 1997), in different situations (Ravald & Grönroos, 1996), pre- and post-purchase (Gardial et al., 1994) and between tangible and intangible offerings (Naumann, 1995). In the

consumer behaviour literature, Rokeach (1973, p. 5) reasoned that ‘a value is an enduring belief that

a specific mode of conduct or end-state of existence is personally or socially preferable to an opposite or converse mode of conduct or end-state of existence’. Various researchers have extended Rokeach’s

work by defining value in terms of mental images and cognitive representations (Peter & Olson, 1987; Wilkie, 1994). In addition, Bhat & Reddy (1998) state that a brand can both satisfy consumers’ practical needs by offering functional value and satisfy consumers’ self-expression needs by providing symbolic value and that one brand can satisfy both values at once. Added value on the other hand is claimed ‘to form the most important part of a brand's definition and separates the brand from the

product’ (Jones, 1986). Grönroos (1997) defines the two concepts in relations to each other in that

11

provide negative added value. Negative added value occurs often in the form of, for example, a poorly functioning mobile application or ill-informed and understaffed customer service. In addition to the underlying meaning of added value, the role of added value was advocated as securing competitive advantage (e.g. Normann & Ramírez, 1994; Naumann, 1995) and long-term success (Naumann, 1995; de Chernatony & McDonald, 1998). Moreover, de Chernatony et al. (2000) point out that the emotional role of adding value to a core offering brings long-term success. Devlin (1998) states especially in the case of more complex service markets it is important to consider which services add value to a company’s offerings as price, for example, is less important compared to simple service markets. For example, for highly intangible service offerings, organisation-wide factors such as the level of functional service quality are perceived important in adding value (Grönroos, 1984, 1987, 1988; Parasuraman et al., 1991). Additionally, other researchers (Zeithaml, 1998; Nayyar, 1990; Bharadwaj et al., 1993) propose extrinsic cues, such as image and reputation, are important when considering value-adding services in cases when consumer understanding of offerings is limited.

In contrast to an abundance of literature on value-adding services, their nature, role and sustainability in different industries (see De Chernatony et al., 2000), literature on value-adding services in the energy market is more difficult to find. Due to its relative infancy in the energy market, not much marketing-oriented research focusing on value-adding services exists. There is, however, a growing interest of these services largely caused by climate change concerns, technology advances and growing consumer involvement (Valocchi et al., 2007). Research in this field has been focused mainly on the technological properties of these services such as smart metering and in-home displays (Darby, 2010). As a result, no well-defined conceptualisations of value-adding services were found. This paper proposes the following definition of value-adding services in its simplest form specifically for the energy market as ‘the provision of insights and advice, by a company to a

customer, based on customer usage data’. More elaborately they can be explained as the provision

12

services could be the reduction of consumers’ electric bills. However, the aim is provide insights for the customer and based on these insights, advice is given to the customer in a two-way

communication stream. This gives it a more elaborate purpose. More concretely, the reduction of energy usage, the creation and contribution to environmental awareness and the increase in customer engagement are all direct goals of these services. The main indirect goal, from a suppliers’ point of view, is increased loyalty.

2.2 Loyalty and value-adding services

The link between value-adding services and loyalty seems apparent. By introducing value-adding services, one aims to increase behavioural loyalty through increasing attitudinal loyalty. As Oliver (1999) discusses, companies in specific situations cannot reach any degree of loyalty whatsoever, often because consumer involvement towards a product or service is too low. One of the conditions to gain loyalty is to provide superiority of your offering thereby achieving differentiation and by introducing additional services this requirement is met. Conversely, the influence of loyalty on the adoption of value-adding services can be argued as well. If customers are satisfied with the product or service they receive from the focal company, they may expand their scope beyond merely one product or service of the company and engage in cross-buying expanding their lifetime value to the firm.

2.2.1 Influence of value-adding services on loyalty

13

available. Options range from no value-adding services to free services providing few functionalities online (little added value) to costly products providing elaborate insights to customers (high level of added value). As a result, the author argues that the higher the level of added value a service provides in addition to core offerings, the more is contribute to the behavioural loyalty of a customer. Hence, based on the discussion above the following hypothesis was developed

H1: The adoption of value-adding services has a positive effect on behavioural loyalty.

As mentioned above, a discussion on behavioural loyalty is incomplete without considering attitudinal loyalty. Reaching behavioural loyalty goes through attitudinal loyalty and is achieved by creating an emotional response towards a brand (Bagozzi, 1992). This emotional reaction drives customer engagement and involvement which drive behaviour. This line of reasoning was also proposed by Mehrabian & Russel (1974) in their Stimulus-Organism-Response (S-O-R) framework. In this framework they claim that when an individual encounters a stimulus (S), (s)he develops internal states (O) which in turn dictate his/her responses (R). Thus, adopting a value-adding service works as a stimulus, evoking an emotional response which affects the attitude towards a company, resulting in higher loyalty in the form of, f.e., cross-buying or contract extension. In the energy market, several studies find evidence for growing customer engagement caused by some of the perks of value-adding services such as energy disaggregation, real time energy information (Chakravarty & Gupta, 2013) and smart metering (Darby, 2010). Hence, positive emotional and attitudinal connections towards the brand are developed. Thus, it is expected that the adoption of value-adding services positively influence attitudinal loyalty. In line with, e.g., Bandyopadhyay & Martell (2007) and the discussion on loyalty, we reason that attitudinal loyalty is an antecedent of behavioural loyalty. Hence, value-adding services influence the purchase of a company’s offerings by a consumer by evoking an emotional response whereby an increased positive attitude is formed. An increased positive attitude towards the brand will result in increased purchasing behaviour. Therefore, the following hypothesis is adopted:

H2: The effect of the adoption of value-adding services on behavioural loyalty is mediated by

attitudinal loyalty.

2.2.2 Influence of loyalty on value-adding services

14

the customer relationship increases the number of services adopted in the banking industry. In addition, if the relationship length between a company and customer increases, the perceived switching costs to switch to another brand opposed to remaining at the current brand increases as well. Also, perceived switching costs increase over time the longer one remains at a company (Jones et al., 2000). Hence, lock-in increases which results in a higher propensity of adopting an additional offering from the same company. Consequently, higher customer lifetime value increases the likelihood of cross-buying. Also, Reinartz & Kumar (2003) found evidence that relationship duration increases spending and cross-buying which increases the propensity that additional products or services are adopted. In addition, a higher degree of behavioural loyalty is expected to lead to the adoption of services which provide a relatively higher level of added value. A higher customer lifetime value is likely to signify higher commitment and trust towards a brand resulting in a higher tendency to broaden the relationship from a customers’ perspective. Therefore, the likelihood of intensifying the adoption of value-adding services and, thus, services compromising a higher degree added value. Hence, the following hypothesis was derived.

H3: Behavioural loyalty has a positive effect on the adoption of value-adding services.

Attitudinal loyalty, similar to behavioural loyalty, has an influence on value-adding services as well. If customers are satisfied with the product or service they receive from the focal company, they may expand their scope beyond merely one product or service of the company and engage in cross-buying. Several studies agree with this as they argue that commitment and trust has to be built up over a longer period of time before a customer broadens the relationship with a company (Kelley & Thibaut; 1978; Verhoef et al., 2001; Reinartz et al., 2008). If one engages in a relationship with a company spending a relative large amount of money or for a long period of time, one is assumed to be satisfied with the products or services provided and is more inclined to share this satisfaction with others. Research by Mägi (2003) supports these notions by finding evidence that satisfaction is associated with an increased share-of-wallet and one way to increase share-of-wallet is cross-buying. Also, attitudinal loyalty affects buying in that customers become more responsive to cross-selling promotions. Consequently, there is incentive enough to believe that attitudinal loyalty increases as relationship length and lifetime value of a customer increase due to the fact that a sustainable relationship was forged and that this stronger attitudinal loyalty enlarges the cross-buying of value-adding services. Therefore, the following hypothesis was derived:

H4: The effect of behavioural loyalty on the adoption of value-adding services is mediated by

15

A bidirectional relationship between loyalty and value-adding services is hypothesised in the past section. Naturally, both directions will not be equal in size and one of the directions will dominate the other. When the dominant direction within the relationship is determined, managers could use these insights to define a starting point and allocate (marketing) resources.

2.3 Control variables

As explained, the energy market deviates from many other industries as it provides a commodity service where demand is dependent on many factors. Therefore, multiple variables are important when considering the energy market. The most relevant factors are described below.

First of all, the type of residence (e.g. (semi-)detached house, terraced house) a customer lives in highly affects the energy usage. Larger, detached houses will logically require more energy to heat. Moreover, often a larger residence size accommodates more people which increases energy usage also. The year of construction of the house affects the energy usage as well, since many houses built more recently are isolated better compared to older houses. Another variable which needs to be controlled for is area of residence. According to GfK (2014), the situation from before the

liberalisation, the dominance of energy providers in specific geographical areas, is still visible. For example, Essent is still dominant in the Eastern part of The Netherlands, whereas Nuon holds a dominant position in the North-West and Eneco dominates the Western and central provinces of The Netherlands. Also, marketing activity may influence the data when not controlled for. One of the consequences of the liberalisation of the energy market is the increase in marketing expenditures (Sullivan & Sheffrin, 2002). Marketing efforts could have a negative impact on the existing customer base. For example, acquisition campaigns may offset existing customers in that they feel betrayed as new customers get additional services or against price reductions (Novo, 2005). In addition,

16

2.4 Conceptual research model

Based on the preceding theoretical discussion, the hypothesised effects between variables are captured in the following conceptual model in figure 1:

17

3. DESIGN

The following chapter discusses the research methodology. First, the data collection procedure and measurement approach will be discussed. Lastly, the sample characteristics and the choice of analyses will be elaborated upon.

3.1 Data collection

For this explorative research, data was attained from two sources: the customer database of Eneco and survey data on consumer socio-demographics by Experian, a global information services group. In terms of data available at Eneco, an effort was made to obtain as many variables as possible. The dataset contains data on 994.231 customers. It consists of variables that will explain constructs such as value-adding services (specific Eneco offerings such as TOON and Mijn Eneco), behavioural loyalty (customer lifetime value) and attitudinal loyalty (satisfaction score). The data compromises cross-sectional data points, more specifically at the end of quarter 2 (June 30th) in 2013 and at the end of quarter 2 in 2015 . Preferably, time series data was used for the analysis, however, it proved

impossible to obtain data over a longer period of time regarding the variable measuring value added services as the product was introduced only recently. The survey data by Experian contains variables that will function as control variables such as type of residence and household size. Both will be discussed in more detail later. Moreover, an attempt was made to acquire usage data from an external partner of Eneco, Qubi, a developer of products and software for the energy market. This appeared to be impossible due confidentiality issues. For multiple reasons, competitor data such as marketing expenditures data and sales data have been excluded as it was unmanageable to obtain these variables within the scope of this research. The data which was acquired will be described in the following section.

3.2 Measurement approach

For each construct it was carefully considered which variables were to be used for the analyses, as were the measurements for each variable. These considerations will be presented below.

3.2.1 Value-adding services

Value-adding services will be measured by means of consumer behaviour of two combined value-adding services labelled Mijn Eneco (ME) and TOON. Mijn Eneco is an online platform allowing customers of Eneco to gain access to personal information, advice and perform actions regarding their energy usage to manage their relationship with Eneco. TOON is a home energy device providing insights in one’s energy usage and costs. Both services provide a certain degree of insights to

18

adoption of a certain service, customer usage measured in log-ins would have been used in the analysis. However due to data availability constraints, it was impossible to acquire this data for TOON. The required data will be extracted from the internal database of Eneco.

3.2.2 Behavioural loyalty

The construct of behavioural loyalty comprehends the customer lifetime value (CLV). CLV measures the economic worth of a customer (Berger & Nasr, 1998). More elaborately, CLV represents value of all current and future profits generated from a customer over the life of his/her business with the firm. Again, the CLV merely measures the economic commitment rather than an attitude towards a brand. This measure is in line with several studies which use current and future revenue streams as a measure of loyalty (Sewell & Brown, 1990; Sasser et al., 1997; Gupta & Zeithaml, 2006). Customers are classified into one of five ordered groups ranging from very low to low, medium, high and very high. The CLV and the comprehending classifications are retrieved from Eneco’s internal database.

3.2.3 Attitudinal loyalty

Due to limitations regarding the type of variable allowed to use for the mediation analysis,

satisfaction was adopted to measure attitudinal loyalty. This construct is measured by the question

‘how satisfied are you with the services of our company on a scale of 0 to 10’. This data will be

collected from an external customer service agency named Xpoints. Ideally, the attitudinal loyalty construct would have consisted of two metrics, namely official Net Promoter Score (NPS) and top-2-box satisfaction. Recent research provides us with evidence that combining these two metrics increases the predictive power of a model (De Haan et al., 2015).

3.2.4 Control variables

Several control variables will be employed. First of all, type of residency consists of the following unordered set: detached house, semi-detached house, corner house, mid-terrace house, apartment, combination and unknown. Area where the customer resides is divided into two categories: within or outside the historically determined geographical area. Building year of residency ranges from 1900 or earlier, 1900-1944, 1945-1969, 1970-1989, 1990-1994, 1995-1999, 2000-2004, 2005 and later, combined to unknown. Moreover, the variables family size (varying from 1 to 7), income (below average, average, 1,5x average, 2 times average, 2,5 times average) and age (ranging from 18-109) are included.

3.3 Model specification

19

behavioural loyalty, mediated by attitudinal loyalty. Secondly, a model has to be developed

describing the effect of behavioural loyalty on value-adding services, mediated by attitudinal loyalty. The two models will be elaborated upon in the following sections.

3.3.1 Model 1

This model estimates the effect of the adoption of value-adding services on behavioural loyalty mediated by attitudinal loyalty. The dependent variable, behavioural loyalty is measured by the customer lifetime value where each customer is allocated to one of the five segments ordered from low to high. As a result, the corresponding dependent variable can be labelled as an ordered

(multinomial) dependent variable. The appropriate model or estimation technique for such a variable would be an ordered logit or probit regression (Franses &Paap, 2001). To present, only few empirical marketing models have been specified as an ordered model (Leeflang et al., 2014), examples are Verhoef et al. (2001) and Sridhar & Srinivasan (2012). An ordered logit model is estimated by Maximum Likelihood estimation methods. It calculates the utility of each customer and scores this customer for each variable. The higher the score, the more likely the customer chooses that

alternative. A logit model would be adopted rather than a probit model as the probabilities are easier to calculate and the use of a logit model is more pragmatic (Aldrich & Nelson, 1984). However, despite the fact an ordered logit best fits the data, an unordered (multinomial) logit is estimated. The main reasons for this will be clarified in the subsequent section. This model will be estimated in two steps. First, a ‘base’ analysis will be conducted to find the effect of value-adding services on

behavioural loyalty without including attitudinal loyalty. As an extension of step 1, an additional analysis will be performed in which attitudinal loyalty will be included as mediation variable in the former specified link.

Based on choice of variables in this and the previous chapter, the following empirical model was developed:

Step 1 (base analysis):

Ui j t-1 = β1j + ∑4j=1β2jVAS i t-1 + β3jAGEi t-1 + β4jUSAGEi t-1 + ∑6j=1β5jHHSIZEi t-1 + ∑5j=1β6jINCOME i t-1 +

∑2j=1β7jREGIONi t-1 + ∑7j=1β8jRESTYPEi + β9jRESYEARi + εi j

Step 2 (mediation analysis): ALi t-1 = β1 + ∑3j=1β2jVASi t-1 + εi j

20 Where:

U = the probability that customer i choses category j of the dependent variable i = customer i = 1,…,N

j = category t = time in years

VAS = value-adding services AL = attitudinal loyalty AGE = age in years

USAGE = electricity usage in t-1 HHSIZE = household size INCOME = income

MOVE = moving out of the house REGION = inside region (0/1) RESTYPE = type of residence

RESYEAR = year of construction of residence ε = error term

3.3.2 Model 2

The second model estimates the effect of behavioural loyalty on the adoption of value-adding services mediated by attitudinal loyalty. The dependent variable, value-adding services, is measured by 4 levels of value that these services can add to core offerings ordered from low to high. Similar to model 1, the dependent variable can be labelled as an ordered (multinomial) dependent variable for which an ordered logit regression analysis is the appropriate model or estimation technique.

Comparable to model 1 an ordered logit is not estimated, rather an unordered (multinomial) logit is estimated. The main reasons for this will be clarified in the subsequent section. Also, this model will be estimated in two steps. First, a ‘base’ analysis will be conducted to find the effect of behavioural loyalty on value-adding services without including attitudinal loyalty. As an extension of step 1, an additional analysis will be performed in which attitudinal loyalty will be included as mediation variable in the former specified link.

21 Step 1 (base analysis):

Ui j t-1 = βj1 + ∑5j=1β2jBL i t-1 + β3jAGEi t-1 + β4jUSAGEi t-1 + ∑6j=1β5jHHSIZEi t-1 + ∑5j=1β6jINCOME i t-1 +

∑2j=1β7jREGIONi t-1 + ∑7j=1β8jRESTYPEi + β9jRESYEARi + εi j

Step 2 (mediation analysis): ALi t-1 = β1 + ∑5j=1β2jBLi t-1 + εi j

Ui j t-1 = βj1 + ∑5j=1β2jBLi t-1 + β3jALi t-1 + β4jAGEi t-1 + β5jUSAGEi t-1 + ∑6j=1β6jHHSIZEi t-1 + ∑5j=1β7jINCOME i t-1

+ ∑2j=1β8jREGIONi t-1 + ∑7j=1β9jRESTYPEi + β10jRESYEARi + εi j

Where:

U = the probability that customer i choses category j of the dependent variable i = customer

t = time in years j = category

BL = behavioural loyalty AL = attitudinal loyalty AGE = age in years

USAGE = electricity usage in t-1 HHSIZE = household size INCOME = income

MOVE = moving out of the house REGION = inside region (0/1) RESTYPE = type of residence

RESYEAR = year of construction of residence ε = error term

22

4. RESULTS

This section will discuss the results from the analyses performed on the data, this will be done separately for both models. Before doing so, some descriptive statistics will be discussed in order to get a first overview of the data.

When one would take a look at the descriptive statistics in table 1 and 2, several issues arise. A few respondents are 109 years old which could indicate a fault in the data. Nonetheless, it is possible that Eneco serves one or more 109 year olds as several people of 109 years or older live in The

Netherlands. Moreover, energy usage shows a relatively high maximum which could possibly indicate an error in the data as well. This could be explained by, for example, a large household size combined with a large and old residence which requires a large amount of energy on a daily basis. Interestingly, the proportion of adopters (for the two highest value-adding services) is relatively low, even though it does increase over time as table 2 shows a higher percentage of adopters. This could be explained by the relative recent introduction of the services resulting in low product awareness.

Table 1: Descriptive statistics at time t-1 (n = 994.231).

Variable Summary statistics

Behavioural loyalty Very low (7.3%), low (20,1%), moderate (38.2%), high (24.6%), very high (9.8%)

Value-adding services No adoption (53.9%), adoption Mijn Eneco (45.2%), adoption TOON (0.1%), adoption

MijnEneco/TOON (0.8%)

Type of residence Detached-house (4.6%), semi-detached house (5.6%), corner house (10.2%), mid-terrace house

(30.7%), apartment (42.2%), combination (0.5%), unknown (6.3%)

Household size 2 (35.6%), 3 (30.2%), 4 (17%), 5 (9.9%), 6 (4.6%), 7 (2.7%)

Income Very low (22.4%), low (26.7%), moderate (25%), high (16.8%), very high (9.1%)

Mean Standard deviation Minimum Maximum

Attitudinal loyalty 5.48 1.118 1 10

Age 54.87 16.87 18 109

Usage 3018.30 2271.312 0 130879

Region 0.91 0.280 0 1

Year of residence 3.40 2.103 1 10

Table 2: Descriptive statistics at time t (n = 994.231).

Variable Summary statistics

Behavioural loyalty Very low (9.9%), low (19.1%), moderate (37.0%), high (23.6%), very high (10.5%)

Value-adding services No adoption (43,3%), adoption Mijn Eneco (5062%), adoption TOON (0.5%), adoption

MijnEneco/TOON (5.6%)

Type of residence Detached-house (4.8%), semi-detached house (5.3%), corner house (8.6%), mid-terrace house

(24.9%), apartment (33.9%), combination (0.5%), unknown (22%)

Household size 2 (34.1%), 3 (30%), 4 (17.5%), 5 (10.5%), 6 (5%), 7 (2.9%)

Income Very low (21.5%), low (26.8%), moderate (25%), high (17.1%), very high (9.5%)

Mean Standard deviation Minimum Maximum

Attitudinal loyalty 7.24 2.008 1 10

Age 52.24 17.31 18 109

Usage 2933 2265 0 134798

Region 0.64 0.48 0 1

23

4.1 Model 1 (2013)

The analysis of this model comprehends two steps. Firstly, a ‘base’ analysis which will test hypothesis 1 followed by a mediation analysis which includes the attitudinal loyalty variable testing hypothesis 2.

4.1.1 Base analysis

By performing the base analysis, it will be tested what the direct effect of value-adding services is on behavioural loyalty. The model assumptions will be checked and model selection will be done before conducting the analysis.

4.1.1.1 Preliminary checks

Before performing an analysis several classification, or decision, trees were created in order to get an easy, comprehendible visualisation of the data. A decision tree provides some initial insights and allows one to observe relationships between variables in a visual manner. Multiple types of decision trees can be created utilising different methods including the CHAID, exhaustive CHAID and CART methods. The reason multiple methods are used and compared is rooted in the fact that if similarities between different trees are found this should give confidence about patterns and

interactions in the data. The trees show a large resemblance in terms of relevant variables as at least 2 out of three trees include the multiple variables thereby suggesting to include these in the analysis. The variables present in each tree are presented in table 7. Somewhat surprising and contrary to literature, attitudinal loyalty is not included in more than one classification tree thereby suggesting it should not be included in the analysis. Additionally, household size and region should excluded in the analysis as well, when relying on the decision trees.

Table 7: Classification trees statistics

Value-adding services Type of residence Household size Income Attitudinal loyalty

Age Usage Region Year of residence Model 1 (CHAID) X X X X X X Model 2 (Exhaustive CHAID) X X X X X X X Model 3 (CART) X X X X X

24

Values greater than 2.5 indicate serious multicollinearity issues in a logistic regression (Allison, 1999). Table 5, in appendix 1, shows that the highest VIF value is 1.326, indicating there are no reasons for concern. Also, when performing an ordered logistic regression analysis, a concern raised by many researchers is the independence of irrelevant alternatives that may be present in a model (e.g. Keane, 1992; Pels et al., 2001). The independence of irrelevant alternatives (IIA) can be described as, all else being equal, a person’s choice between two alternative outcomes is unaffected by what other choices are available to him (Cheng & Long, 2007). The logit models assume all alternatives are independent which might cause problems when two alternatives are in fact not independent but similar. As a result, the odds of the different categories relative to each other may change. Therefore, one needs to test for this and one method of doing so is to perform multiple analyses and each time exclude one category from the dependent variable. If the parameter estimates differ significantly, the assumption is violated and a solution (in the form of a nested logit or probit analysis) would be needed. The outcomes of the analyses when excluding the different categories compared to the ‘original’ model, showed only minuscule differences. Consequently, the IIA assumptions are not violated allowing the author to proceed righteously.

As mentioned in the previous section, an ordered multinomial logistic regression analysis will be conducted to obtain parameter estimates for our model. Several models were estimated which included different variables, partly based on the outcomes of the decision trees presented above. This was done in order to find the best performing combination of variables in a model. The outcome of the tests proved to be significant across all models in terms of goodness-of-fit and likelihood ratio test. This means that the model fits the data and that the models with variables are significantly better than the intercept only model. This allows one to correctly interpret the parameter estimates. Despite the fact that these assumptions are met another, rather important, assumption was violated. When conducting an ordered multinomial logit, it is assumed that the parameters of different

categories of the dependent variable are the same. This is called the assumption of parallel lines and when the assumption does not hold, it means there is no parallelity between categories (Erkan & Yildiz, 2014). In this instance, the null hypothesis, stating that the slope coefficients in the model are the same across response categories, is rejected thereby concluding that the ordered logit

25

test. Another option is to eliminate one or more non-critical variables from the analysis. Again, the fewer slopes to estimate, the lower the propensity that the test assumptions would be violated. This was applied by excluding several variables (age, usage, household size, residence size, residence age, region and income) simultaneously or separately, from the analysis. This did not result in an

increased significance of the parallel lines test either. The last solution to overcome a failed test is to switch to a multinomial logit regression analysis. The downside of this solution is that one loses information on the ordering of the dependent variable. Rather than the dependent variable consisting of several categories which were ordered along the same continuum, the categories are now separate entities and will be assessed accordingly. When, for example, a variable consists of three categories, each category will be assessed separately in relation to the other two rather than being assessed simultaneously with all three categories. Despite the fact that one loses information when conducting a multinomial logit model rather than an ordinal logit model, this method will be used in this paper.

4.1.1.2 Model selection

In order to obtain the most optimal outcomes from the analysis, a ‘best’ model needs to be selected. It is common to produce multiple models with the same model specification and compare them with each other to select the best performing one (Leeflang et al., 2014). An important condition of model comparison is that all models should hold the assumptions discussed before. This is the case,

therefore, the models are allowed to be compared. When comparing different models, several criteria to assess the model fit of a model can be employed of which the most widely used ones are the Bayesian Information Criterion (BIC), the pseudo R2 and likelihood ratio (LR), Hit Rate (Leeflang et al., 2014). The BIC is based on an information theoretic rationale and the precision of the estimated model can be represented by the natural logarithm of the likelihood L. The lower the values for BIC, the better the model fit. Another criterion for assessing the model fit is a pseudo R2. For many other regression tests other than a logistic regression, R2 is the correct statistical measure to assess the model fit. That is because, logistic regression uses the maximum likelihood estimates (MLE) rather than an ordinary least squares (OLS) statistic (Zumbo & Ochieng, 2002). The pseudo R2’s are based on comparing the log-likelihood of a model with only an intercept (LL0) with the log-likelihood of a

model with K explanatory variables (LLk). In this paper, the Nagelkerke R2 is used, mainly because it

corrects for the number of observations. The higher the value of this measure the better the model fit. An additional criterion to assess the model fit is the hit rate which is determined by the

26

likelihood ratio which compares a model with parameters to an intercept only model, this is

comparable to an ANOVA F-test. To compare the models with each other the requirement is that one model is a nested one of the other (Leeflang et al., 2014).

Table 8 below shows the outcomes of the different criteria for different models. The ‘base’ model 1 includes all available variables, namely behavioural loyalty, residence type, residence age, region, household size, age, income and usage. Model 2 excludes the variables age, usage and income and model 3 excludes all variables except for the independent variable behavioural loyalty. First of all, the assumptions of all three models, as described above, were satisfied indicating we can safely compare the different models. As table 8 shows, model 1 scores best on the Nagelkerke R2 and the Hit Rate and model 3 scores best on the AIC, BIC and likelihood ratio. Based on these numbers, one could say either model 1 or model 3 is to be chosen for the analysis. Model 1 would provide the best goodness-of-fit whereas model 3 provides us with the simplest, most parsimonious model in line with Little (1970). Nonetheless, model 2 is selected. The reason for the selection of model 2 lies in the fact that goodness-of-fit and model complexity must be balanced against each other (Myung, 2000). First of all, a model primarily based on goodness-of-fit will result in a model which will fit the data really well but is overly complex, resulting in the model will generalising poorly to other data (Hertz et al., 1991; Bishop, 1995; Ripley, 1996; Myung, 2000). As this research is explorative of nature, it is increasingly important that the results must be generalisable across other data, in order to contribute to existing and future literature. On the other hand, purely basing a model on the BIC will result in an extremely simple model. As the BIC corrects for sample size (making it a more suitable criterion than the AIC which does not account for this) and the sample size in this paper is relatively large, the BIC is optimal in a model with only one independent variable. There is evidence that in practical

applications, the BIC tends to select models which under fit the data (Burnham & Anderson, 2002). Thus, in order to find a model which is well-balanced in terms of goodness-of-fit and

parsimoniousness, model 2 was chosen. The variables which were included in model 2 were selected based on how much each variable would contribute to the measures of Nagelkerke pseudo R2 and BIC. The variables which were excluded decreased the Nagelkerke pseudo R2 the most and BIC the least.

Table 8: Model comparison*

AIC BIC Nagelkerke Hit Rate Likelihood Ratio Model 1

(all variables)

1837913.205 1838974.499 .383 52.5% 1837729 (p < .01) Model 2 (excluding age,

usage, income

19559.028 20314.850 .230 46.6% 19431.028 (p < .01) Model 3 (excluding all

variables except for behavioural loyalty)

206.235 395.190 .069 41.6% 174.235

27 4.1.1.3 Parameter interpretation

Table 9 shows the parameters estimates derived from the multinomial logistic regression analysis. The estimates should be interpreted differently from other regression analyses such as a linear analysis. The parameter estimates for a multinomial logit, exp(B), have to interpreted as being odds or ‘relative risk’ to the ‘base’ case. In other words, the odds ratio for a unit increase in an

independent variable for category x relative to the base case given that all other variables in the model are held constant. Each variable has its own ‘base’ case indicated by ‘.b’ meaning, for example, that all behavioural loyalty categories are compared to the category ‘very high’. To begin with, all results which are being discussed are highly significant (p ≤ .01) if not indicated otherwise. Thus, from table 9 it can be derived that for category ‘TOON’ relative to ‘TOON/Mijn Eneco’, the odds of having a CLV value which is ‘very low’ relative to having a value of ‘very high’ would be expected to increase by a factor of 1.721 given all other variables in the model are held constant. In other words,

customers who adopted ‘TOON’ are more likely than customers who adopted ‘TOON/Mijn Eneco’ to have a CLV of ‘very low’ rather than a CLV value of ‘very high’ by a factor 1.721. More specifically, the chance is 72.1% larger for customers in the category ‘TOON’ as opposed to category ‘TOON/Mijn

Eneco’ to be of ‘very low’ value rather than of ‘very high’ value. In addition, the remaining levels of

value-adding services can be interpreted as follows: customers in the value-adding services

categories ‘Mijn Eneco’ and ‘no value-adding service’ are more likely to be of ‘very low’ lifetime value than customers who have adopted ‘TOON/Mijn Eneco’ rather than being of ‘very high’ lifetime value by a factor 3.390 and 1.870, respectively. With regard to the CLV categories ‘low’, ‘moderate’ and ‘high’, similar results are observed. Relative to customers who adopted ‘TOON/Mijn Eneco’, customers who adopted ‘TOON’, ‘Mijn Eneco’ and ‘no services’ are more likely to be in the CLV category ‘very low’ compared to ‘very high’ by a factor 1.721, 3.390 and 1.870, respectively. Also, relative to customers who adopted ‘TOON/Mijn Eneco’, customers who adopted ‘TOON’, ‘Mijn

Eneco’ and ‘no services’ are more likely to be in the CLV category ‘low’ compared to ‘very high’ by a

factor 1.420, 7.021 and 5.712, respectively. Additionally, relative to customers who adopted

‘TOON/Mijn Eneco’, customers who adopted ‘TOON’, ‘Mijn Eneco’ and ‘no services’ are more likely to

28

Table 9: Parameters estimatesac

. The reference category is 5. b

. This category is set to zero because it is redundant.

c

. The complete output of the analysis can be found in table 16 in appendix 2.

Behavioural loyalty B Std. Error Wald Df Significance Exp(B) 95% Confidence interval

for Exp(B) Very low behavioural loyalty

Intercept -3.684 .041 8160.034 1 .000 . . Value-adding services 0 .626 .019 1117.698 1 .000 1.870 1.803 _1.940 Value-adding services 1 1.221 .017 4971.464 1 .000 3.390 3.277 3.507 Value-adding services 2 .543 .054 100.015 1 .000 1.721 1.547 1.914 Value-adding services 3 0b . . 0 . . . .

Low behavioural loyalty

Intercept -3.663 .032 12794.880 1 .000 . . . Value-adding services 0 1.743 .017 10178.382 1 .000 5.712 5.522 5.909 Value-adding services 1 1.949 .017 13544.485 1 .000 7.021 6.794 7.255 Value-adding services 2 .351 .057 37.696 1 .000 1.420 1.270 1.589 Value-adding services 3 0b . . 0 . . . .

Moderate behavioural loyalty Intercept -3.947 .027 21551.440 1 .000 Value-adding services 0 2.549 .015 29691.181 1 .000 12.792 12.426 13.168 Value-adding services 1 2.523 .014 30909.279 1 .000 12.469 12.123 12.825 Value-adding services 2 .399 .049 65.859 1 .000 1.491 1.354 1.642 Value-adding services 3 0b . . 0 . . . .

High behavioural loyalty

29

4.1.2 Mediation analysis

By conducting the mediation analysis, the effect of attitudinal loyalty in the link between behavioural loyalty and value-adding services is established. In mediation analysis, several relationships exist, as depicted in figure 2. Mediation assumes several relationships between variables. First of all, the direct effect of the predictor (X) predicting the dependent variable (Y). Secondly, the predictor (X) predicts the moderator variable (Z). Thirdly, the moderator (Z) predicts the dependent variable (Y). Combining the last two effects comprehends the mediation effect.

For the assessment of the mediation effect, Baron & Kenny (1986) developed 4 conditions which must hold before a mediation effect is recognised. The conditions are as follows: (1) Effect of X on Y (C) has to be significant, (2) Effect of X on Z (A) has to be significant (3) Effect of Z on Y (B) has to be significant, (4) Z has to influence the effect of X on Y (C’) as depicted in figure 2. Not only do these effects need to be significant, they also need to be substantially different from zero. If either one of the conditions is not met, the mediator variable does not possess any mediating power. This analysis was performed by using the MEDIATE macro developed by Hayes (2012). For the analysis, a

bootstrapping method was used. Bootstrapping is a method that involves repeated sampling, 20,000 times in this analysis, from the dataset and estimates the indirect effect of ab in each resampled data set. The reason to perform a bootstrap is that the small sample size (N = 250) could be too small for straightforward statistical inference making and this problem is solved with bootstrapping. Following Preacher & Hayes (2008), a 95% confidence interval was used for the parameter estimates.

In the mediation analysis, the variables used are different compared to the variables used in the ‘base’ analysis. The reason for this is that the MEDIATE macro developed by Hayes (2012) requires the dependent (Y) and moderator (Z) variable to be continuous in order for them to be interpreted properly. As a result, the dependent variable, behavioural loyalty, was altered into a continuous