“ Horses for Courses? ” : The influence of customer activity on the effectiveness of Relationship Marketing Instruments

(1)

“ Horses for Courses? ” : The influence of customer activity on the

effectiveness of Relationship Marketing Instruments

Michael Thomas Sants

University of Groningen

Faculty of Economics and Business

MSc Marketing (Marketing Intelligence Track)

MSc Thesis

June 2020

Michael Thomas Sants S2490293

Visserstraat 19 9712CR Groningen m.t.sants@student.rug.nl

First Supervisor: R.P. (Roelof) Hars MA, MSc

Second Supervisor: Dr. E. (Evert) de Haan University of Groningen Faculty of Economics and Business

(2)

(3)

Management Summary

The field of Customer Retention Management (CRM) is a vital component in every successful business strategy. The customer journey plays its part in here as it allows

companies to follow the activities that customer undertake during their relationship with the firm. Calls in the literature have been made regarding a more refined view of CRM which is where this thesis contributes. One of the instruments of CRM is Relationship Marketing, which is employed in most day-to-day business operations. Customers are targeted with numerous Relationship Marketing Instruments (RMI) such as Cross Sell Attempts, Deep Sell Attempts, Free Benefits and Social Relationship Marketing practices. However, there are discrepancies within the literature regarding the effect of these Relationship Marketing Instruments. Some authors have stated these discrepancies can be attributed to the difference in customer activity which results in certain RMIs not being effective for every type of customer which is what this thesis investigates.

(4)

Acknowledgements

To whomever it may concern,

If you are reading this, you are reading the final output of my MSc Marketing at the

University of Groningen. Starting with a BSc in International Business in 2013, an exchange semester in China and countless other adventures now we are here. The past year has been a tremendous learning experience for which I would like to express my gratitude towards the full team of (Assistant) Professors who I have encountered during my time at the MSc Marketing. The journey to this point was insightful and I am grateful for being able to have taken part in such an adventure. Moreover, I would like to thank my first supervisor Roelof Hars for his accurate feedback and efforts undertaken in assessing my preliminary works.

Furthermore, before I commence on an extension of my academic journey at BI Oslo, I want to take this moment to express my gratitude towards my parents and sister for always fully supporting me and giving me the ability to pursue my dreams. Next to this, I want to thank my friends for their support and especially Dawid with whom I enjoyed this tremendous learning experience the past year.

Michael T. Sants June 2020

(5)

3.2 Variable Selection ... 24 3.3 Outlier Analysis ... 27 3.3 Choice of Technique ... 28 3.3 Selection Bias ... 29 4. Model Selection ... 33 4.1 Model Comparison ... 33 4.2 Multicollinearity ... 35 4.3 Plan of Analysis ... 36 5. Hypotheses Testing ... 36 5.1 Model Estimation ... 36 5.2 Further Assessment ... 44 6. Discussion ... 46 6.1 Managerial Implications ... 48

7. Limitations and Future Research ... 49

8. References ... 51

Appendices ... 59

Appendix 1: Distribution of IVs per Customer Type ... 59

Appendix 1.1: Customer Type A... 59

Appendix 1.2: Customer Type B ... 61

... 61

Appendix 1.3: Customer Type C ... 62

(6)

Appendix 2: Out-of-Sample Validity Measures per Customer Type... 64

Appendix 2.3: Customer Type C ... 66

Appendix 2.4: Customer Type D ... 67

Appendix 3: Tables of Descriptive Statistics per Customer Type ... 69

Appendix 3.3 : Customer Type C ... 71

Appendix 3.4: Customer Type D ... 72

Appendix 4: Full Model Estimation per Customer Type ... 73

Appendix 4.1 Customer Type A Model Estimation ... 73

Appendix 4.2 Customer Type B Model Estimation ... 74

Appendix 4.3 Customer Type C Model Estimation ... 75

Appendix 4.4 Customer Type D Model Estimation ... 76

Appendix 5: Further Assessment of relation between activity and Relationship Marketing ... 77

Appendix 5.1: Model Specification ... 77

Appendix 5.2: Model Estimation... 78

(7)

Introduction

Managing customer attrition has been a key challenge in customer relationship management (CRM) (e.g. Blattberg, Kim and Neslin, 2008; Reinartz, Krafft and Hoyer, 2004; Reinartz and Kumar 2000). Recent papers have established the ever-increasing need within service industries to manage customer attrition (Ascarza, Iyengar and Schleiger 2016; Wieringa & Verhoef, 2007). Next to this, the need to manage customer attrition is gaining relevance due to increased information availability which in turn results in an increased ease in opportunity to switch to a competitor (Jahromi, Stakhovych & Erwing, 2013). Consequently, customer retainment is a key response to prevent such attrition. According to Reinartz, Thomas & Kumar (2005), retaining a customer is more beneficial for the company’s profits compared to acquiring new customers. This is confirmed by Ibáñez, Hartmann & Calvo (2006) who state that new customer acquisition costs could potentially be greater than the cost of customer retainment. To illustrate, according to Blattberg et al. (2008) annual churn rates can be as high as over 60%, which highlights the importance of the field of customer retention. Furthermore, companies that focus on prioritizing the prevention of customer attrition are better off compared to firms who emphasize lowering acquisition costs or increasing profit margins when it comes to firm value (Gupta, Lehman and Stuart, 2004). Next to this, according to Buchanan and Gilles (1990), retained customers are more cost-efficient since they are aware of the business processes in place. Furthermore, the authors raise the point of decreased customer attrition having an easing effect on the day-to-day work environment of employees which contributes to overall organizational efficiency (Buchanan and Gilles, 1990).

To be able to cope with the phenomenon of churn (i.e. customer attrition), organizations try to instate churn prediction models in industries where customers have contracts such as the service industry (Verbeke, Dejaeger, Martens & Hur, 2012). Therefore, identification of customers prone to switching carry high priority (Keaveney & Parthasarathy, 2001). As a result of accurate churn prediction, marketing resources can be utilized more efficiently (Neslin, Gupta, Kamakura & Lu 2006). In current customer retention literature, there is an emphasis on churn prediction, with approximately 41,900 papers on Google Scholar alone that have the words ‘churn prediction’ in the title.

(8)

2018). To illustrate, Ascarza et al. (2018) raise the point that there is a need within the literature to identify when and how a customer can be retained. An estimated 49% of managers are unsatisfied with their retention capabilities, indicating a practical need for further understanding on how to effectively retain customers (Forbes Insights 2014). Furthermore, in a survey conducted by Handley (2013), 85% of the customers questioned indicated that their service provider could undertake more efforts to try and retain these customers. This illustrates that next to further understanding, which customers churn, there is an increasing need to understand on how to retain customers. In order to effectively identify the customers with the highest probability to churn, it is essential to comprehend when people churn and what effects marketing efforts have on preventing churn (Verbeke et al., 2012). Furthermore, Bolton, Lemon and Verhoef (2004) and Berry (1995) propose a split between (relationship) marketing instruments (RMIs), where on one side of the continuum these RMIs can be of a transactional character, whereas on the other side they can be of a more relational character. This continuum will be delineated later on in this thesis. Next to this, not every consumer is interested in building a relationship with a firm, which results in relationship marketing not always being a preferable strategy since not all consumers are as prone to engaging in this relationship (Berry 1995; Crosby, Evans and Cowles 1990; Sheth and Parvatiyar 1995). Furthermore, Berry (1995), Barnes (1997) and Anderson and Narus (1998) all argue that service organizations may need to pursue strategies other than a relational one. To illustrate, Barnes (1997) argued that there are instances where customers do not want a relation with their service provider and thus engaging in RM might exert a negative effect. The above-mentioned discrepancies regarding RMI applicability have thus been pertinent in literature since the previous century.

(9)

retention had a positive effect on decreasing the number of customers who churned. Therefore, it can be stated that there is disagreement within the literature regarding the effectiveness of relationship marketing strategies and its effect on customer retention, which is where this thesis contributes.

The before-mentioned disagreement is characterized by the difference in the outcomes of relationship marketing efforts conducted. Two examples of extreme outcomes are highlighted; on the one end of the spectrum there is Ascarza et al. (2016), who approaches customers with an improved price plan, whereas on the other end there is Burez & van den Poel (2007) who target customers with free items (e.g. free movie tickets, invitation to an event). In the 2016 study by Ascarza et al., the authors find an increase in customer churn as the result of employing this form of RM whereas Burez & van den Poel (2007) find an increase in customer loyalty. Both of these forms of Relationship Marketing can be attributed to the Type I sphere as stated by Berry (1995) and Bolton et al. (2004). This split will be delineated later in this thesis. However, although both actions are positioned within the transactional side of Relationship Marketing (Berry (1995), Bolton et al,. (2004)), there are discrepancies in their effectiveness. Therefore, this thesis posits that the generalization within this framework might not be applicable to every customer. A potential factor aligning the contradictory findings is the level of customer activeness of the customer targeted. In the paper by Burez & van den Poel (2007) the customers selected were those who voluntarily and actively communicated their phone number with the company and thus displayed a form of increased activeness (actively providing company with their phone number) whereas in the field experiment conducted by Ascarza et al. (2016), the authors promptly lower customer inertia by randomly targeting (passive) customers.

(10)

Consequently, the following questions arise, which this thesis aims at answering:

- ‘What are the effects of different types of relationship marketing instruments have on a customer’s churn probability?’

- ‘What effect does a customer’s level of activeness have on the effect of relationship marketing instruments effectiveness with regards to the churn probability?’

Combining the above-mentioned sub-questions results in the following main research question: ‘What are the effects of different relationship marketing instruments on a customer’s churn probability and how are they influenced by a customer’s level of activeness’

The remainder of this thesis is structured in the following way : The upcoming section displays a theoretical framework, which discusses the relevant literature. Furthermore, within this part, the hypotheses are formed, and a conceptual model is derived. Afterwards, the research design is outlined. Following this, the dataset is going to be explained and thereafter the hypotheses testing methods incorporated throughout this thesis will be presented. Consequently, the results are going to be discussed and its managerial implications will be highlighted. Furthermore, potential recommendations for future research will be presented and any limitations that arise are going to be considered.

Theoretical framework

Churn

(11)

preventing, since the firm could potentially control this, whereas external/involuntary churn is uncontrollable from a firm’s standpoint. As specified prior, a firm has an influence on the customer’s churn probability. Since one of the aims of this thesis is to deliver applicable managerial implications, the focus within this thesis will be on voluntary internal churn since this could aid managers in developing an effective marketing strategy without influencing the churn probability of the targeted customer in a negative way. Furthermore, some CRM literature uses intention to churn (e.g. Wieringa et al. (2007)) whereas throughout this study the binary churn definition as stated by Neslin et al. (2006) is incorporated and therefore follows actual churn behavior rather than intentions to churn.

Relationship Marketing

(12)

Relationship Marketing Instruments (RMIs) are marketing instruments that influence the success of customer relationships (Verhoef, 2004; De Wulf, Odekerken-Schröder, and Iacobucci 2001; Bhattacharya and Bolton 2000). RMIs are those type of programs that emphasize the developing of customer’s relationships, and thus extend beyond the direct marketing promotions such as telemarketing and coupons (Bolton, Lemon and Verhoef, 2004).

According to Bhattachary and Bolton (2000), RMIs are a subgroup of marketing tools specifically targeted at facilitating the mutual cooperation between the customer and firm. Next to this, Berry (1995) distinguishes between two levels of relationship marketing. At the first level, which is labeled Type I RMIs, the instruments are characterized by economic efforts where the emphasis is on the transactional nature of the relationship. These can be classified on whether there is an economic incentive at play for the client, related to the transactional sphere of the relationship. Examples of this are price recommendations (Ascarza et al., 2016), cross-sell attempts (Kamakura, 2007) and the obtainment of free benefits with economic value (Burez & van der Poel, 2007) plus any other RMI effort undertaken in the transactional sphere in the relationship. Due to the transactional character of the RMI effort, scholars have also labeled this as more product focused (e.g.; Bolton et al.,2004, Berry, 1995).

The second level (Type II) of RMIs, have an emphasis on social attributes (Bolton et al., 2004). These Type II RMIs are characterized by providing social benefits to the customer (Bhattacharya and Bolton, 2000; Dabholkar, 1994). By employing the Type II RMIs, the firm aims at giving the customer a more personalized feeling by integrating social attributes (Verhoef, 2004). The type II RMIs are more in the relational sphere rather than transactional, as their Type I counterparts are (Berry, 1995). Examples of this, in the social sphere, are intangible benefits, unrelated to the transaction at hand, that positively influence the relationship by increasing affection (Bolton et al., 2004). A broader overview is going to be provided later in the literature review.

(13)

maintained their relationship with the company behave as advocators for the organization by spreading their experience through word-of-mouth (Zhang et al., 2010; Evans et al., 2001). Moreover, long-term customers are more profitable compared to their short-term counterparts (Van der Poel and Lariviere, 2004). The rationale for engaging in RM efforts is thus a clear one according to these scholars. Next to this, a majority of research and practice assumes that relationship marketing strategies result in an enhanced relationship which creates a positive effect on the firm’s performance (Crosby, Evan and Cowles 1990; Morgan and Hunt, 1994). There has been research conducted that disagrees with the aforementioned predominantly rosy view of RM. According to other scholars in the field this is not the case and the effectiveness of relationship marketing strategies can be labeled as ineffective and potentially having a negative effect (De Wulf et al., 2001; Hibbard et al. 2001; Colgate and Danaher, 2000). Furthermore, Moorman et al. (1992) argue that high levels of interaction between a customer and firm could trigger a belief that one can no longer objectively evaluate the importance of the relationship which exerts a negative effect. Furthermore, Gruen et al. (2000) argue that customers may develop a ‘What have you done for me lately’ attitude if there has been too little interaction.

Hibbard et al. (2001) argue for a more refined view on RMIs which is supported by the more recent call in order to identify how and when to target customers by Ascarza et al. (2018). Furthermore, Ascarza et al. (2018) state that there are several aspects of the managerial side of customer retention that have not received an adequate amount of attention of which one is a more refined view of customer retention efforts as an organizational tool. This thesis will commence with highlighting the distinctions within RM and their expected different effects by scholars in order to attempt to contribute to the desired refined view of the managerial side of customer retention in the form of RMI.

Type I RMIs: Transactional Sphere

(14)

of the relationship. Examples of this are price recommendations (Ascarza et al., 2016), cross-sell attempts (Kamakura, 2007) and any other RMI effort undertaken in the transactional sphere in the relationship. Due to the transactional character of the RMI effort, scholars have also labeled this as more product focused (e.g.; Berry, 1995).

A prominent example of a type I RMI activity is a cross-sell attempt. Cross-selling relates to the presentation of tailored offerings in order to broaden the assortment of services/products a client already has with the firm (Malms & Schmitz, 2011). Cross-Selling has transormed into a RM strategy by broadening the scope of the relationship and thus getting the customer more involved (Kamakura, 2007). Furthermore, by broadening this scope, it is argued that it increases the costs of switching for the customer. These switching costs refer to both monetary and psychological and thus aid in customer retention (Kamakura, 2003). However, cross-selling lacks the personal character that is usually pertinent in RMIs efforts and could thus have a negative effect in the form of the customer churning (Kamakura, 2007). In relation to this, Kamakura (2007) states that cross-selling can be perceived by the customer as too intense and thus relates to the view of Grayson and Ambler (1999) that cross-selling can feel like the customer’s trust in this relationship is being exploited. This potential misalignment of effectiveness could stem from the applicability of a type I RMI to a certain type of customer and its willingness to extend or alter the relationship with the service provider. As stated by Barnes (1997), not every type of customer is willing to build a relationship in general and as showed by Ascarza et al. (2016) potentially targeting inert customers could result in a negative effect.

(15)

Another example of a transactional relationship marketing instrument is a deep-sell attempt. This form of transactional RM refers to altering the relationship between a customer and firm in terms of changing the plan/subscription that a client currently has with a company but within the context of the current product. An example of this would be a price recommendation or recommending a different plan for the same product, where there is a new contract developed for this service however the type of service technically remains the same. Ascarza et al. (2016) argue that first of all that by recommending a different plan, customer inertia is lifted in their field experiment and thus could potentially cause the customer to switch. However, there are instances within this field experiment where the price recommendation does not cause churn which therefore results in conclusion that the effect of the RMI is thought to be highly customer dependent. This again infers that a deep-sell attempt is not ineffective per se, however, for this type of transactional relationship marketing might not be effective for every type of customer.

(16)

Although effectiveness for these RMIs is thus customer dependent according to the literature, one thing they all have in common is that they are in the transactional sphere. Furthermore, selling techniques such as cross selling and deep selling refer to remaining within the current relationship with the firm. It relates to altering or broadening the current relationship with the service provider. As stated by Barnes (1997) and Berry (1995) this might not be attractive to certain customers, which in turn can be caused by difference in customer inertia as stated by Ascarza et al. (2016).

Combined with the above-mentioned misalignment of effectiveness, of both examples of a transactional form of RM, the hypothesis can be drawn that the effectiveness of a transactional RMI could potentially be dependent on the nature and inertia of the customer. Furthermore, it becomes evident that some customers are simply not willing to build a relationship with their service provider. These aforementioned factors are hypothesized to have an influence, which this thesis will further investigate in the forthcoming chapters. Consequently, hypotheses will be drawn regarding the effects of transactional RMIs and the influence of customer activity.

Type II RMIs: Relational Sphere

The second level (Type II) of RMIs, have an emphasis on social attributes (Verhoef, 2004). These Type II RMIs are characterized by providing social benefits to the customer (Bhattacharya and Bolton 2000; Dabholkar, Johnston and Cathey, 1994). By employing the Type II RMIs, the firm aims at giving the customer a more personalized feeling by integrating social attributes (Verhoef, 2004). The type II RMIs are more in the relational sphere rather than transactional, as their type I counterparts are (Berry, 1995). These are characterized by emphasizing the social character of the relationship by sending for example personalized emails (Verhoef, 2004).

(17)

wishing a happy birthday, holiday emails or a newsletter are example of social relationship marketing efforts based on the criteria that they enhance the personal relationship between firm and customer (e.g. Berry, 1995).

These social marketing efforts lead to an increased level of the perception that preferential treatment is at place for a customer. Peterson (1995) argues that personalized treatment enables a company to touch upon a customer’s basic human need to feel important. Prior literature on service marketing identifies and recognizes the influence of interpersonal interaction on customer satisfaction. For example, Brown, Fisk & Bitner (1994) state that within service settings, the happiness of a customer is influenced by the extent and quality of the interaction between the customer and the service provider. Furthermore, if firms are able to create a hedonic experience, such as increased affection which yields hedonic utilities to the consumer, this is hypothesized to induce positive feelings for a customer (Bhattacharya and Bolton 2000; Muniz and O’Guinn 2001).

(18)

drawn that the effectiveness of a relational RMI is thus dependent on the nature of the customer which this thesis will further investigate in the forthcoming chapters. Consequently, hypotheses will be drawn regarding the effects of relational RMIs.

Customer Activity

One of the potential factors driving the discrepancies within the effectiveness of RMIs on certain customers could be customer inertia (Ascarza et al., 2016). Customer inertia is identified as a continuous attachment which is to a large extent not driven by emotions (Gounaris and Stathakopoulos, 2004; Lee and Cunningham, 2001). Inert customers steer clear from learning new service routines and avoid comparing different prices (Pitta et al. 2006). According to Wieringa and Verhoef (2007), inert customers keep purchasing from the same supplier and as the result of inertia diverge from explicit decisions which results in habitual buying behaviors. Recalling the claim of Ascarza et al. (2016) that targeting inert customers results in an increased churn probability, customer activity could be a potential indicator of RMI effectiveness. Furthermore, Wieringa and Verhoef (2007) state that inertia can be captured in econometric marketing models by including past behavior which is where this thesis commences.

Van Doorn et al. (2010) state that customer engagement behaviors result from motivational drivers. Furthermore, Roos & Gustafson (2011) have done research on what triggers have influence on active and inactive customer behavior. Throughout their research, they have defined active customers as ‘active customers are those who search for information in order to be able to make a deliberate and conscious decision.’. In contrast, an inactive customer is defined as ‘passive customers are those who have not searched for information and therefore has fewer conscious reasons for his or her decision’. The definition by Roos and Gustafson (2011) is in line with the differentiation between inertia within customers as proposed by e.g. Pitta et al. (2006) where an inert customer is defined as a customer who steers clear from learning new service routines, searching for information and avoids comparing different prices. The difference in level of activity could thus be an indicator of customer inertia according to the above-mentioned authors.

(19)

process of the customer. In other words, the core activities relate to internal customer activity. These activities are under direct influence of the service provider and can be communications sent to the customer, the customer browsing the service provider’s website or any other interaction between the customer and the service provider and can thus be attributed to the relational sphere. Furthermore, Grönroos and Voima (2012) label this as the joint sphere of value creation. In this dyad, both the firm and the customer contribute to the co-creation of utility for a customer. Therefore, this form of activity can be labelled in the current relational sphere due to both parties participating in this dyad. Consequently, one can infer that if a customer does not participate in this joint sphere of value creation as stated by Grönroos and Voima (2012), this customer is less interested in engaging with their focal service provider which is not an uncommon occurrence since some people simply do not want to build a relationship with their service provider (Barnes, 1997).

Another form of customer activity, which is (usually) invisible to the service provider is labeled by Mickelson (2013) as related customer activity. Although usually unobservable to the firm, for this thesis there is unique data regarding this phenomenon is available and provides an interesting opportunity to retrieve insights. For example, the data contains information regarding contact with competitors of the focal service provider, e.g. comparing prices or exploring different alternatives. Even though outside of the perception of the focal service provider, these activities are vital to the customer’s value creation process (Grönroos and Voima (2012)). They are vital to the value creation process in the sense that these related customer activities aid in a customer’s value creation by engaging in a different dyad than which they are currently in. In this context, value is created by engaging in a potential other relationship with a competitor or at least exploring the opportunity (Mickelson, 2013). Combining both types of activities, where a customer can thus be inactive and active in either one of the activity spheres, the following framework can be created:

Table 1: Customer Activity Framework

With regards to the effectiveness of the RMIs discussed earlier in this literature review, it became evident that its effectiveness is highly dependent on the nature of the customer,

(20)

potentially a customer’s inertia and thus level of activeness. As stated above in Table 1, internal activity is of a dyadic form where both parties contribute to create value for the relationship and thus attributes to the relational sphere. Next to this, external activity is outside of the scope of the focal firm and refers to the transactional sphere in which customers are exploring potential other service providers.

Recalling the properties of the Type I and Type II RMIs where a Type I RMI can be labelled as the transactional form of relationship marketing and the Type II RMIs are characterized as the relational form of relationship marketing. Therefore, in order to attempt to align the paradoxical findings of the papers regarding relationship marketing as discussed above, the aim is to find out whether there is a specific type of RMI that works for a specific type of customer. As the key difference between the papers of Burez & van den Poel (2007) and Ascarza et al. (2016) are the types of RMI and the level of customer inertia, taking into account the type of RMI and the level of customer activity may aid in aligning these contradictory findings of RMI effectiveness. Furthermore, due to some customers simply not being interested in building a relationship with their service provider (Barnes, 1997) this supports the notice that there might be certain actions suitable for only a certain type of customer. As internal activity refers to co-value creation within the relational dyad (Grönroos and Voima, 2012, Mickelson, 2013) it can be hypothesized that if a customer is internally active there is at least some form of willingness to build a relationship with the service provider (Mickelson, 2013). Contrary, if a customer is externally active, which in the scope of this thesis refers to exploring competitive options, there is at least some willingness present to explore options outside of the current relational dyad (Mickelson, 2013). Furthermore, it can be inferred that the higher the level of activity internally, the more structural the co-value creation is which in turn increases commitment in the relationship (Grönroos and Voima, 2012).

(21)

and relationally oriented, where the two are thus intertwined. Based on this, the following can be hypothesized:

H1: ‘Engaging in Type I RMIs has a negative effect on a Type A customer’s churn probability’ H2: ‘Engaging in Type II RMIs has a negative effect on a Type A customer’s churn probability’

Next to this, a Type B customer is characterized by being internally inactive thus less interested in co-value creation within the current relationship (Grönroos and Voima, 2012). Contrary, this type of customer is active in the external sphere and thus rather transactional oriented as stated by Mickelson (2013). Consequently, the RMIs in the transactional sphere might appeal more to this type of customer whereas this customer could be less interested in the relational type of RM efforts due to its increased interest in the current relationship. Ergo, the following can be hypothesized:

H3: ‘Engaging in Type I RMIs has a negative effect on a Type B customer’s churn probability’ H4: ‘Engaging in Type II RMIs has a positivee effect on a Type B customer’s churn probability’

Furthermore, a Type C customer is internally active which indicated that the customer is relationship focused in the sphere of co-value creation as stated by Grönroos and Voima (2012). Contrary, this type of customer is externally inactive which indicates that the transactional aspect might matter less to this customer. Therefore, the following can be hypothesized:

H5: ‘Engaging in Type I RMIs has a positive effect on a Type C customer’s churn probability’ H6: ‘Engaging in Type II RMIs has a negative effect on a Type C customer’s churn probability’

Lastly, the final type of customer, is a Type D customer. This customer can be labelled as completely inert displayed by the inactivity in both spheres of customer activity. As stated by e.g. Asczarza et al. (2016) and Wieringa & Verhoef (2007), approaching inert customers can result in an increased churn probability or increased intention to churn. Therefore, the following can be hypothesized:

(22)

Control Variables:

Cooil et al (2007) call for the need to account for customer heterogeneity since certain differences are to be expected. This call is supported in a churn-based context by e.g. Wieringa and Verhoef (2007). Therefore, certain social demographics are going to be considered within this thesis as potential covariates, of which one of them is Age. According to Lambert-Prandaud et al. (2005), older customers are less prone to switching and thus have a lower churn probability compared to their younger counterparts. According to Lambert-Prandaud et al. (2005) this can be attributed to older people having an increased aversion to change and an increase in cognitive decline. Therefore, it can be stated that age is thought to have an effect on a customer’s churn probability. This is supported by Svendsen and Prebensen (2013), who state that younger customers are more likely to churn.

(23)

Conceptual Model

Figure 1: Conceptual Model

3. Research Design

This chapter will go into detail of the data used, how it is tidied for analysis and will be described. Outliers and oddities will be discussed together with missing values. Furthermore, the model specification chapter will be highlighted here together with a model comparison in order to see which model is most suitable for analysis. Afterwards, within the model selection effects will be accounted for in order to ensure an unbiased sample.

3.1 Data Collection

The dataset used throughout this thesis entails a dataset of a European financial service provider. The primary dataset ranges from 1st of June 2015 until 29th of April 2019 and contains

the data of 15.000 customers who were still a client of the firm at the first of June in 2015. To ensure anonymity of the firm, as requested, I will not go into further detail here. Moreover, I will try to touch upon the specific elements in the dataset that form the variables created in a manner that ensures discretion as requested by the European financial service provider whilst trying to provide clarity for the reader.

3.1.2 Initial Data Cleaning

(24)

filtered out in order to ensure that only those who churned voluntary are being considered within the analysis. Furthermore, in order to create compatibility for merging different (sub)datasets there has been ensured that spelling mistakes have been taken care off (e.g. Septembr changed to September) and that all dates within the datasets are set to the format of YYYY-MM-DD. These logical inconsistencies were taken care of in order to avoid potential problems.

3.2 Variable Selection

In order to assess the effect of different marketing action on a customer’s churn probability, marketing campaigns of A European financial service provider were grouped together in one overarching variable to reflect the RMI a specific customer has received. Therefore, cross sell attempts initiated by the firm were grouped together and named Cross Sell, deep sell attempts were grouped together and labeled Deep Sell. A similar approach was taken for Free Benefits and Social RMIs. A threshold was set at 1000 customers reached in order to ensure a scale large enough to allow for effective interpretation (e.g. seeing a difference in churn with relation to this specific campaign). However, not every customer has an equal relationship length with the firm and may not have received as many marketing campaigns compared so some of their counterparts due to their relationship length. Therefore, a variable was created called ‘Relationship Length’which reflects the relationship length in years. This variable was used to modify the implied overarching RMI variables which results in RMI predictors in which the relationship length of a customer with the firm is accounted for. The same principle is applied to both internal and external activity since also here, not every customer has an equal relationship length with the firm and thus this was taken into account.

The following part of the thesis provides an overview of the predictor variables utilized within this research. The variables considered in this section are the variables as displayed in the conceptual model, Figure 1.

(25)

Cross Sell: This variable is the grand total of a variety of cross sell attempts received by a customer during their time of being a client at the firm. The choice for an overarching cross sell results from aiming at parsimony with the to-be specified model (Little, 1993). Therefore, the variable cross sell can be interpreted as a numeric variable and thus translates to the total amount of cross sell attempts received. In order to keep the model parsimonious, a variety of cross sell attempts have been summed together into the variable Cross Sell in order to reflect the total amount of cross sell attempts received by a customer during their time of being a customer at the company.

Deep Sell: The variable Deep Sell is the total sum of a variety of deep sell attempts that a customer has received during their duration of being a customer at the company. This variable is of numerical nature. In order to keep the model parsimonious, these deep sell attempts have been grouped together into the variable Deep Sell in order to reflect the total amount of deep sell attempts a customer has received.

Free Benefits: This variable is the total count of free benefits received. Different free benefits have been received by customers and due to the requested anonymity of the European financial service provider this thesis will not list each unique case. However, these benefits can be found in the spheres of for example free tickets to unique events. These are thus free benefits outside of the scope of the transaction at hand between the service provider and the customer. Furthermore, in order to ensure parsimony, free benefits received by the customer are summed up which results in the variable Free Benefits. This variable reflects the total count of free benefits received by the customers whilst being a customer at the European financial service provider.

(26)

Regarding customer activity, Ascarza et al. (2017) call for the need to experiment with including fine-grained data such as customers visiting certain web pages in order to broaden the knowledge in the area of customer retention management. The reason for its potential positive effect in including this type of data is that these data have a high dimensionality and extreme sparsity (de Fortuny, Enric, Provost, 2013). Therefore, a similar approach will be adopted in quantifying internal and external customer activity which the data available allows for.

Internal Customer Activity: Internal Customer Activity is quantified here as the total individual count of internal webpages a customer has viewed during their relationship with the European service financial provider. The examples here are in line with what is found in the literature (e.g. Grönroos & Voima (2012), Mickelson (2013). They refer to webpages viewed regarding the relationship at hand. This can be certain details or other information regarding the relationship with the focal service provider. This is a numerical variable, where the more pages viewed relates to the customer being more active internally.

External Customer Activity: External customer activity is quantified as the total number of price comparisons a customer has made on a certain comparison website (following the line of reasoning in e,g.; Mickelson, 2013). This is a numerical variable, where the more pages viewed relates to the customer being more externally active and thus has viewed more price comparisons regarding a similar product as what they have now with a financial European service provider.

Age: The dataset used for this thesis also entails some demographics which are implemented as control variables within this research. One of these relevant social demographics of the customer is age. This variable Age is the exact age of the customer at the beginning of their relationship with the company. Age is a numerical variable. In order to be able to interpret the effect of age on a customer’s churn probability, with regards to the intercept, the variable is mean centered in order for the intercept to not reflect a hypothetical customer with the age of zero.

(27)

When this variable takes the value of 0, it relates to a male customer whereas the value of 1 refers to the customer being of the female gender.

3.3 Outlier Analysis

Further exploratory analysis prior to analysis is required. The above-listed variables in section 3.2 will be inspected for outliers. An assessment of the values that could be classified as outliers was made by employing boxplots. The overview of the boxplots can be found in Appendix 1. Furthermore, it can be inferred from the Appendix 1.1-1.4 that most predictor variables showed values that are traditionally considered as outliers as specified by the values falling outside of the whisker range of the boxplots. However, a majority of these were deemed nonproblematic since they were deemed realistic employing face validity. Furthermore, a boxplot considers a value an outlier based on percentiles and therefore when a large number of observations has the same value, the boxplot can thus falsely bias this as an outlier. This is the case for most of the variables since they can take values of zero due to for example not receiving a specific RMI action. Therefore, combining both face validity and statistical theory regarding the outlier classifications by means of a boxplot, the outliers detected in most variables were deemed realistic and thus not removed.

However, for the variable external activity this was a different case. As visualized in Appendix 1.1 and 1.2 in Appendix X, one can infer that for a big part of these values, they deviate extremely from the rest of the data (e.g. observations of 3000 pages viewed on a yearly basis). Due to the comparatively low number of observations for this variable, compared to the other variables at hand, the decision was made to truncate the variable rather than remove outliers that deviated strongly in order to sustain the number of observations and thus aid statistical power. The truncating value was set at 500 pages viewed on a yearly basis. Robustness checks were performed in terms of comparison, where values of e.g. 1000 or 750 pages resulted in certain values of the predictor variable perfectly predicting customer churn. Therefore, the decision was made to truncate the variable Customer External Activity at 500 external webpages viewed rather than remove the values causing the distorted analysis due to the already relatively low number of observations at hand.

(28)

distribution is derived from a real life distribution due to it deriving from company data, this is acknowledged and continued with. Regarding the different customer segments, it becomes clear that there are certain demographical differences at play in terms of age. The inert segment has an average age of 57 which is in line with the common notice that older people are less technologically equipped and thus have a decreased ability to be active. Furthermore, the only externally active segment is the youngest with an average age of 43. The rest of the dispersion of values seems similar which can be found in Appendix 2.

3.3 Choice of Technique

The dependent variable of interest throughout this thesis is churn, which is classified as a binary variable. This variable can take two values; either zero (does not churn) or one (customer churns). Following Leeflang et al. (2015), if a marketing problem comes with a binary dependent variable, this requires a binary logit or probit model. For both of these models, there are two discrete outcomes; 0 or 1. An estimated logit or probit model does not focus on estimating observed values but instead estimates probabilities (Leeflang et al., 2015). To make sure that the estimated probabilities are within the range of possible outcomes (e.g. not smaller than 0, not over 1), either a logit model (derived from a logistic cumulative distribution) or a probit model (derived from a standard normal cumulative distribution) is necessary. Applying one of these two methods results in the estimated probabilities falling within the desired range (Leeflang et al., 2015). Both types of models are thus justified for this thesis, however, as the result of mathematical convenience the logit model is often preferred in marketing applications (Leeflang et al. 2015). Therefore, this thesis proceeds with adopting a binary logit model.

According to Fok (2017), it is assumed that there is an observed, latent variable Ui that drives the binary decision, in this case churn of individual i. This latent variable can be interpreted as the (indirect) utility of staying with the company. The utility has the following linear specification (Fok, 2017):

(29)

Where the error term is part of the utility that the customer is aware of. Therefore, from the point of view of the customer equation 2 is not uncertain. However, in this form of analysis the 𝜀_𝑖 is not observed and therefore the decision is treated as a random variable. As a result of this, a Cumulative Distribution Function (CDF) is applied to the binary logistic regression, displayed in equation 3:

The densities of the logistic distribution are symmetric around zero. Therefore, for these distributions it holds that F(z) = 1- F(-z) for all values of z (Fok, 2017). Taking this into account the probability of observing a one (churning) in the logit model can be written as:

Where Pr [Yi = 1] shows the probability of a customer churning or not, where the logical consistency requirement of the probability taking a value between zero and one is satisfied (Fok, 2017). Furthermore, 𝛼 is the intercept, x’i is a vector of the explanatory variable, 𝛽 is a vector of the coefficients of xi..

3.3 Selection Bias

According to Bolton and Lemon (1999) and Verhoef, Bolton, Lemon (2004) certain customers are selected to receive campaigns and promotions that have the highest response probability. Companies do this using statistical techniques such as chi- squared automatic interaction detection or CHAID. Therefore, in order to be able to try and effectively draw causal interferences from models estimated throughout this thesis there needs to be accounted for selection effects.

(30)

due to the fact that this type of method relies on including observable confounds. Contrary, for example a two-step sample selection model such as the well-known Heckman Model relies on including unobservable confounds due to the incidentally truncated nature of the depend variable in that equation (Heckman, 2013). Furthermore, the Heckman correction involves a normality assumption (Heckman, 2013), which is not met due to the binary nature of the dependent variable in this thesis. Next to this, propensity scores can be estimated with observed customer characteristics which resembles a practical implication of utilizing available unique company data. Due the availability of unique customer characteristics data for this thesis, propensity scores are preferred. Based on the above-mentioned reasons, it is decided to continue with Propensity Score Methods in order to account for selection bias within this thesis.

Propensity scores can be utilized to decrease the selection bias at hand in four ways , namely; Propensity score matching (Rosenbaum and Rubin, 1983), Inverse Probability of Treatment Weighting using the propensity score (Morgan and Todd, 2008), Stratification based on the propensity score (Abadie, Drukker, Herr, Imbens, 2004; Luncefor and Davidian, 2004) and Regression Adjustment using the propensity score (Elze et al., 2017).

A potential limitation of employing propensity scores for matching, which is another well- known technique for eliminating selection bias, is that a considerable number of subjects is needed in the control group (Guo and Fraser, 2015). Consequently, depending on the different matching technique, and the number of subjects being matched to either the control – or treatment group, a large number of those subjects assigned may not be used (Randolph and Falbe, 2014). Therefore, potentially data is lost and consequently the overall analyses are negatively affected. Given that in any research setting data collection is costly for both types of subjects, one should prefer techniques that incorporate all the observations (Abadie and Imbens, 2016). Therefore, it is chosen not to employ propensity scores for matching in order to eliminate selection bias.

(31)

scores could cause biased estimates. Furthermore, applying stratification based on propensity scores potentially results in marginal performance due to the creation of a large set of strata and may not account for strong confounding compared to for example regression adjustment, as has been confirmed by Elze et al. (2017). Next to this, a potential limitation of adjusting a regression with propensity scores is that this is the only propensity score method that requires the creation of a regression model modelling the outcome to the treatment status and the propensity score as a covariate (Austin, 2001). Since the creation of the model mentioned above is already implied (See section 3.3; Model Specification) this limitation is thus not relevant for this thesis and positions the method of regression adjustment as the preferred one in order to account for selection bias

Therefore, based on the above-mentioned limitations, this thesis proceeds with applying regression adjustment based on propensity scores as the method to eliminate selection bias. It has to be noted, that the model estimated that results in the propensity scores, needs to specify correctly in order to properly eliminate the selection bias at play (Austin, 2001). As defined by Rosenbaum and Rubin (1983) a propensity score is derived, which thus reflects the conditional probability of being assigned to the treatment group (Selected for Marketing Action) taking into account a set of observed covariates and can be displayed mathematically as (Rosenbaum and Rubin, 1983):

(32)

marketing action) and 11053 can be assigned to the treatment condition (e.g. did receive a marketing action).

The next step is to identify covariates which could potentially affect a firm’s decision on whether to target a customer or not. One can infer that relationship length is a factor of interest (e.g. Rossi et al. (1996), however, Malthouse and Blattberg (2005) conclude that due to the difficulty of accurately assessing Customer Lifetime Value, there are other factors at play compared to solely incorporating relationship’s length. Here a data-driven approach is taken, based on the available company data. First of all, socio-demographic variables were modeled in the form of a binary logit model in order to find out whether they had a significant influence. The binary dependent variable in this case is whether the customer is part of the treatment group or not (e.g. received at least one of the relevant marketing actions or not). Eventually 12 socio-demographic predictors had a significant influence on the probability of a customer receiving a marketing action or not. These covariates cannot be explained further in detail because of requested anonymity by the European financial service provider but are all of socio-demographic nature and are for example income, style of living and geographic location. Next to the socio-demographic variables, relationship length in years also had a signficiant influence and was thus included in the model.

After the model is specified and estimated, the model used was checked for multicollinearity. Regarding the presence of multicollinearity, this has been tested for using the Variance Inflation Factor (VIF) score statistic. These VIF scores are computed by investigating the extent to which one of the independent variables are able to be expressed as a linear regression of the other predictor variables (Leeflang et al. 2015). A VIF over the value of five is perceived as problematic (De Vaus, 2013; Leeflang et al. 2015). There were no VIF values detected that surpassed the threshold of 5, therefore it can be deemed that multicollinearity was not an issue in the final model. After model estimation, the corresponding propensity scores were computed and added to our working dataframe in order to be added as a covariate to the models used for estimation.

(33)

4. Model Selection

4.1 Model Comparison

In order to find the model most suitable for analysis, statistical tests were performed in order to aid in determining the suitability with the data for each model. First of all, a loglikelihood ratio test was This is outcome is displayed by the χ2 statistic. If the loglikelihood ratio test is

significant at a P-Value of <0.05, the direct effects model is better compared to its null model counterpart and therefore adding direct effects is justified. Furthermore, as a robustness check to justify the assumption that a logit model is most applicable, the logit models are also compared to their probit counterparts in order to see if a logit model has the most appropriate fit for the data at hand. Furthermore, each model for each corresponding customer type is compared to its counterpart which has the propensity scores added in order to see if adding the propensity scores to the model is justified, e.g.; aids in enhancing model fit and thus adds in explanatory power. Next to this, the choice between a parsimonious model (where all the different types of RMI I discussed are aggregated together) or a full model (where they are specified in the model individually) is evaluated. In Table 2 below, CTA refers to Customer Type A, CTB refers to Customer Type B and so forth.

Table 2: Model Comparison: In Sample Validity Measures

AIC BIC

McFadden

Pseudo R2:

Cox & Snell Nagelkerke

LR χ2

Logit Models

Model CTA Par 3139.72 3181.79 0.2206 0.2544 0.3458 Sig*

Model CTA Full 3100.89 3154.99 0.2313 0.2649 0.3601 Sig**

Model CTA Full PPS 3088.54 3148.65 0.2349 0.2685 0.3649 Sig***

Model CTB Par 920.06 948.06 0.1621 0.2003 0.2677 Sig*

Model CTB Full 904.03 941.37 0.1806 0.2204 0.2946 Sig**

Model CTB Full PPS 899.78 6986.08 0.1864 0.2266 0.3029 Sig***

Model CTC Par 6999.97 7041.60 0.1419 0.1407 0.2143 Sig*

Model CTC Full 6930.58 6986.08 0.1509 0.1490 0.2269 Sig**

Model CTC Full PPS 6907.76 6970.20 0.1539 0.1517 0.2311 Sig***

Model CTD Par 4600.18 4630.94 0.0353 0.0459 0.0624 Sig*

Model CTD Full 4527.89 4571.17 0.0513 0.0659 0.0897 Sig**

(34)

Probit Models

Model CTA Full PPS 3153.71 3180.93 0.2336 0.2672 0.3632 Insig****

Model CTB Full PPS 902.45 944.46 0.1836 0.2237 0.2990 Insig****

Model CTC Full PPS 7059.05 7107.62 0.1444 0.1430 0.2178 Insig****

Model CTD Full PPS 4593.88 4630.99 0.0523 0.0672 0.0915 Insig****

*= Significant at P <0.01 compared to the corresponding null-model **= Significant at P<0.01 compared to the corresponding parsimonious model *** = Significant at P<0.01 compared to the corresponding full model without Propensity Scores

**** = Insignificant to full logit model with Propensity Scores

In order to assess which versions of the models are most applicable, different validity measures will be used. In regression modelling, it is practice assessing the association between the

dependent variable and the independent variable in a model by looking at the R2 . However,

logistic regression is not suitable for this type of assessment hence the pseudo- R2measures are

_𝑖

Model One: Full Binary Logit Model With Propensity Scores

Where: (definitions can be found in Section 3.2) 𝑈_𝑖 = Probability of Churning

CS = Cross Sell DS = Deep Sell FB = Free Benefits SRMI = Social RMI IA = Internal Activity EA = External Activity A = Age

G = Gender

(35)

Where for different customer types, the model is a nested version of Model One with the activity variables included according to whether a customer is active in that sphere or not. A correctly specified model should be simple, complete, adaptive and robust (Little, 1993). Following the model specification prerequisites in Leeflang et. al (2015), simplicity was achieved by grouping certain business operations into overarching marketing variables which capture the essence of the original individual business operations (e.g. multiple cross-sell attempts grouped into the variable cross-sell). Furthermore, the model is complete in the sense in that it not only encompasses the effect of relationship marketing on its own but captures the effect of different forms of relationship marketing. Next to this, it’s adaptiveness stems from the fact that other potential variables that might exert an effect can be included in the model whilst still achieving the other criteria applied to a correctly specified model according to Leeflang et al. (2015). Lastly, its robustness is achieved by displaying correct marginal effects and changes therein. This means, following the train of thought by Leeflang et al. (2015), that each marginal effect is plausible over an extensive range of potential values for each predictor variable (this relates mostly to an appropriate functional form). Therefore, one can infer that the model is correctly specified following the criteria by Little (1993) and Leeflang et al. (2015).

Next to this, model performance is taken into account in order to confirm that the model selection in Table 2 is the right one. According to Greene and Milne (2010), model performance is measured in practice by assessing the Top Decile Lift, (TDL), GINI Coefficient and the Hit Rate. In terms of assessment, a higher score on these out-of-sample validity measures means a higher model performance and the corresponding statistics for the models displayed in Table 2 can be found in Appendix 2. Inferring from these statistics, it becomes clear that the models picked in Table 2 also display the highest TDL, GINI Coefficient and hit rate and thus this thesis commences with estimating those models since both on in- and out-of-sample validity measures they have the best relative scores.

4.2 Multicollinearity

(36)

surpassed the threshold of 5, therefore it can be deemed that multicollinearity was not an issue in the final model.

4.3 Plan of Analysis

In order to assess the effect of Type I and II RMIs on the different types of customers, four separate models are estimated. The reason for not pooling the data, even though the models are nested versions of one another, is that it is hypothesized that certain spheres of activity (e.g. active and inactive internally/externally) are only relevant for certain groups of customer (e.g. only active/inactive internally, or active/inactive externally) in which a subsampling approach is favored over including an interaction effect alone (Clogg, Petkova and Haritou ,1995). Therefore, to estimate the sole effect of RMIs on activity a subsampling approach is favored. After estimation, to infer whether the differences are significant, an interaction approach is taken as proposed by Clogg, Petkove and Haritou (1995).

5. Hypotheses Testing

5.1 Model Estimation

This section aims at estimating the models specified in section 4.1. For each hypothesized customer type the corresponding model is going to be estimated and its output will be highlighted below. The output resulting from the estimation of all four models(with propensity score matching as a covariate) are presented in Appendix 5.1-5.4. According to Fok (2017), there are three manners of interpreting the output of a (binary) logistic regression namely: Beta Coefficients (Estimates), Odds Ratios and Marginal Effects. In order to be interpretable, the value must be of statistical significance as indicated by the column P-Value. However, insignificant values will still be reported since this could also deliver interesting insights (e.g; something was expected to be of significant influence but was not). Please note: due to the potential confounds within this study, the estimates should not be interpreted causally. There are always other factors at play influencing the estimates, hence this warning prior to interpretation. To continue, the complete output of Model One for Customer Type A, B, C and D can be found below in Appendix 5.1-5.4.

(37)

(significant) value exerts a negative effect on the probability of a customer churning (i.e. Y =1) when there is an increase in that variable. Next to this, the odds ratios offer a similar interpretation to the beta coefficients since odds ratios are solely the exponent of the beta coefficients and thus display similar effects, just in different compositions. To illustrate, if the odds-ratios exert a value of above 1, a positive (increasing) effect on a customer’s churn probability (i.e. Y=1), whereas when the value is below one an opposite effect is in place where there is a decreasing effect on a customers’ churn probability at play. If the Odds Ratio is exactly 1, it can be inferred there is no relationship between the predictor variable in question and the probability of a customer churning.

However, the marginal effects allow for a different interpretation of the logit model and are recommended in practice to enhance an efficient interpretation of results. Therefore, the effect of different RMIs on the churn probability for the different customer types will be assessed using this metric. Table 3 below visualizes the marginal effects of Model One for each customer type.

Table 3: Table of Marginal Effects

Customer Type A Customer Type B Customer Type C Customer Type D Intercept - - - - Type I RMIs Deep Sell 0.1077** 0.1239* 0.0526*** 0.1077*** Cross Sell -0.0369** -0.0704* -0.0367*** -0.0369** Free Benefits 0.0352 -0.0229 0.0038 0.0352*** Type II RMI Social RMI 0.0154*** 0.0108*** 0.0145*** 0.0154***

Internal Customer Activity 0.0087** - 0.0042 -

External Customer Activity 0.0190** 0.0038** - -

Control Variables

Age (Mean Centered) -0.0083** -0.0016 -0.0042 -0.0083

(38)

A negative (positive) marginal effect can be interpreted as a decrease (increase) in the probability of a customer churning given an instantaneous increase in the predictor variable. For marginal effects, caution is advised, since marginal effects can differ for other points in the data since their relationship is of a logarithmic nature and not linear. Therefore, the marginal effects reported assume that all the other covariates displayed are at their average level. Furthermore, in this thesis, this is on a yearly basis. To illustrate, a change of x in the number of Deep Sell attempts a customer of type A would receive in a year, consequently leads to an increase of 0.1077x in the probability of that specific customer churning (Ceteris Paribus). In other words, if this customer receives one additional deep sell attempt the churn probability of this customer increases with 0.1077 whereas for two deep sell attempts this amounts to 0.2154 etcetera. (Ceteris Paribus). It can thus be inferred that a deep sell attempts exerts a

The column labeled estimates displays the beta-coefficients. If this estimate takes a positive (significant) value, an increase in that variable exerts an increasing effect on the probability of a customer churning (i.e. Y =1), whereas a negative (significant) value exerts a negative effect on the probability of a customer churning (i.e. Y =1) when there is an increase in that variable. Next to this, the odds ratios offer a similar interpretation to the beta coefficients since odds ratios are solely the exponent of the beta coefficients and thus display similar effects, just in different compositions. To illustrate, if the odds-ratios exert a value of above 1, a positive (increasing) effect on a customer’s churn probability (i.e. Y=1), whereas when the value is below one an opposite effect is in place where there is a decreasing effect on a customers’ churn probability at play. If the Odds Ratio is exactly 1, it can be inferred there is no relationship between the predictor variable in question and the probability of a customer churning.

(39)

receives one additional deep sell attempt the churn probability of this customer increases with 0.1061 whereas for two deep sell attempts this amounts to 0.2122 etcetera. (Ceteris Paribus). It can thus be inferred that a deep sell attempts exerts a potentially increasing effect on a customer’s churn probability if all other covariates are held at their average value. When a cross sell is employed as an RMI for this type of customer, an additional cross sell attempt received decreases this customer’s churn probability with 0.0369 thus exerting a potential negative effect for the average customer (Ceteris Paribus). Regarding this type of customer receiving free benefits, this estimate is statistically insignificant and thus it’s marginal effects cannot be interpreted. Furthermore, if a customer of this type receives an additional social RMI effort the churn probability of this customer increases with 0.0154 with every additional social RMI unit received thus displaying a potentially increasing effect on customer’s churn probability.

This type of customer is active in both spheres of activity hypothesized, regarding internal activity, with every additional webpage viewed this customer’s churn probability increases with 0.0087 per additional webpage viewed (Ceteris Paribus) whereas for external customer activity the probability of this type of customer churning increases with 0.019 per additional external webpage viewed. Regarding the control variables, if a customer’s age increases with one additional year a customer’s churn probability decreases with 0.0083 per additional year (Ceteris Paribus). Furthermore, the effect of gender on a customer’s churn probability of the A type is deemed insignificant and thus its marginal effects cannot be interpreted.

“ Horses for Courses? ” : The influence of customer activity on the effectiveness of Relationship Marketing Instruments