The effectiveness of a multi-vendor loyalty program on customer behavioural loyalty

(1)

The effectiveness of a multi-vendor loyalty program

on customer behavioural loyalty

Does it really work?

BY ESTHER MENTING

(2)

2

The effectiveness of a multi-vendor loyalty program

on customer behavioural loyalty

Does it really work?

BY ESTHER MENTING 15-02-2017 S2802937 Korreweg 250, 9715 AP Groningen esthermenting1993@hotmail.com 0683796658

Master thesis Marketing Management & Marketing Intelligence

University of Groningen Faculty of Economics & Business

Department of Marketing PO Box 800, 9700 AV Groningen

(3)

3 ABSTRACT

Already some research is done regarding the effectiveness of multi-vendor loyalty programs (MVLPs) on customer behavioural loyalty. This study extents the current literature by investigating into more detail the underlying stages in changing customer behaviour. This is done in two ways: 1) measuring customer behavioural loyalty as both the purchase probability and the purchase size, 2) dividing the type of collaboration into two involvement levels: exchange partner, and exchange and save partner. These different collaboration types go along with different rewarding situations. Where a customer can only redeem points when the company is an exchange partner, it can redeem and collect points when the company becomes both an exchange and save partner. The findings reveal these rewarding aspects of the MVLP (i.e. point redemption and point collection) do positively influence the customer’s buying probability and thus positively influence the number of customers per month. These effects are stronger when a company becomes more involved with the MVLP. On the other hand, the MVLP does negatively influence the size of a purchase and thus the monthly spending. Despite the negative effect on monthly spending, customers who are involved in the MVLP do generate a higher revenue due to their increased buying probability. Overall, the result indicate that customers who are involved within the MVLP are more loyal customers, but might also be more price sensitive.

(4)

4 PREFACE

“The true method of knowledge is experiment (William Blake)”

(5)

5

1. INTRODUCTION

Nowadays the consumer market is far from stable: consumer’s trust is declining and the number of bankruptcies among businesses is rising (CBS 2016). This situation has created an environment where companies are in fierce competition for customers. Over the years, multiple strategies are developed to acquire and/or retain customers. Some companies use pricing strategies (Danzing, Hadar and Morwitz 2014), while others make use of loyalty programs (LPs) (Sharp and Sharp 1997). In 1981, the first LP was introduced by American Airlines. In the subsequent decades businesses became more customer-centric focused, resulting in a growth in the number of LPs (Dorotic, Bijmolt and Verhoef 2012). This grow continued and its impact became visible during 2000s. Twenty years after the introduction of the first LP, 92% of the customers in the United Kingdom were active in a LP (Berman 2006). Recent research showed that approximately 85% of global consumers participate in at least one loyalty program in 2015 (Aimia Inc. 2015). But the interesting question which arises is whether LPs do positively contribute to the acquisition and retention of customers.

The growing interest in LPs is not only shared among managers for business purposes, but also within academic research. Many researchers have studied this field. Some studies investigated the characteristics of LP (Berman 2006; Rese, Hundertmark, Schimmelpfennig 2013; Yi and Jeon, 2003), while others already researched the effects of LP (Arranz and Cillán 2006; Beck, Chapman and Palmatier 2015; Gómez, Leenheer, Heerde, Bijmolt and Smidts 2007; Leenheer, Bijmolt, Heerde and Smidts 2002; Lewis 2004; Sharp and Sharp 1997). Although a lot of research is done on the results of LPs, contradicting findings are presented. Most researchers argue that LPs do have a small positive effect, but some found a negative effect on the spending of the customers. Sharp and Sharp (1997) argue that the effect of such programs deviate per design. Since the findings regarding the effect of LPs are contradicting and/or could vary across program and business, it seems most justified for a business to investigate whether their LP has a positive effect based on their own data.

(8)

8

before, a distinction is made between the different kind of collaborations within a MVLP when investigating the effect on customer loyalty behaviour. This research extends the current literature by making a distinction in the collaboration types, namely: exchange partners (partners at which a customer can only redeem their points) and both exchange and save partners (partners at which a customer can both redeem and collect their points). This distinction is very interesting for both MVLP and current/potential partners of the MVLP. It gives insights into the changing behaviour of the partner’s customers and how members of the program do value the different components of the MVLP (redemption of and collecting points), as a result of different levels of involvement of the partner with the program.

In order to gain insights into the effect of a MVLP on customer loyalty behaviour, real life data form two databases is used. First, data is extracted from the system of a Dutch MVLP, namely XXX. XXX is a loyalty program which is active in The Netherlands, exploited by XXX. Unlike a lot of companies, their LP is not an additional service to the customers, it is their business. The network of XXX consists of multiple partners, in different market segments. The second dataset which is used for this research comes from one of these partners, namely ***. This airline company became in August 2014 a member of the XXX loyalty program. *** started as an exchange partner. In May 2016 the company became besides an exchange partner also a save partner.

1.2 PROBLEM STATEMENT

Like every business, XXX wants to acquire new partners and retain their current partners. In order do this, the company is continuously trying to update, maintain and renew their network of partners. When acquiring a new partner, XXX has to convince the potential partner that their program and network contributes to the performance of the new partner, i.e. XXX has to prove that the MVLP has a positive effect on their customers. Since the literature is not unanimous about the effect of LPs, research is needed to understand the effects. Therefore research is done regarding the effect of the MVLP on one specific partner, namely ***. During this research the next main research question will be investigated: ‘What are the effects of a co-operation with a multi-vendor loyalty program on the customer behavioural loyalty of its partners?’.

(9)

9

A. ‘What is the influence of becoming partner of a MVLP (both exchange, and exchange and save partner) on the monthly spending of its partner’s customers?’

B. ‘What is the influence of becoming partner of a MVLP (both exchange, and exchange and save partner) on the number of customers per month of its partner’s customers?’ C. ‘What is the influence of redemption behaviour on behavioural loyalty?’

D. ‘What is the influence of redemption behaviour on behavioural loyalty, when the value of a point is increased?’

E. ‘What is the influence of the collection of points on behavioural loyalty?’

1.3 CONTRIBUTION

This research extends the current since it is the first research which takes the relationship between the firm and the MVLP into account, by making a distinction between the two different collaboration levels (exchange partner vs. exchange and save partner). Previous literature has already investigated the effect of a MVLP on certain customer outcomes, but never looked at the influence of different types of collaboration on customer outcomes. This distinction is very interesting, since it gives insights into the question whether the different types of collaboration do have varying effects on customer behavioural loyalty. This answer is interesting for both MVLP and current/potential partners of the MVLP. First of all, it gives useful insights for the program, since it provides information about how the members of the program do value the different components (i.e. is redemption a trigger of behavioural loyalty, or is it both redemption and collecting points). But on the other hand, provides it information for current/potential partners about the profitability of the program, and which collaboration type is most gainful. Besides, making this distinction contributes to the current understanding of how LPs do influence customers in their behaviour, i.e. do customers change their behaviour towards a company more positive, resulting in behavioural loyalty, when the company is more involved in the MVLP.

1.4 OUTLINE

(10)

10

2. THEORECTICAL FRAMEWORK

Nowadays, businesses are trying to create loyalty among their customers (Yi and La 2004). The focus on loyal customers gained importance over time, since it is proven that these customers are more beneficial for companies (Reichheld and Teal 1996). Reichheld and Teal (1996) do describe a couple of benefits of loyal customers: higher growth in revenue per customer, higher commitment, lower risk in switching probabilities and more willing to pay a premium price. Based on these potential benefits, it is obvious that a business wants loyal customers. But before a company can benefit, loyalty must be created among its customers. One way to enhance this is through the use of a multi-vendor loyalty program (MVLP). This research tries to investigate the effects of a MVLP on behavioural loyalty. Besides this main effect, the moderating effect of personalized marketing and redemption behaviour of the customers is tested. In this review all four elements of this research are examined: behavioural loyalty, (multi-vendor)loyalty programs, personalized marketing and redemption behaviour. This chapter ends with an overview of the expected relationship between those elements in a conceptual model.

2.2 CUSTOMER BEHAVIOURAL LOYALTY

(11)

11

When investigating the effect of a LP on the actual customer behaviour, behavioural loyalty is the most relevant component to measure since it measures the outcome which entails the effect. Therefore this research focuses on this type of loyalty to address the change in behaviour. According to Sharp and Sharp (1997) behavioural loyalty can occur in six different types: 1) lower switching probability to non-program brands, 2) increased purchases/(start) buying at the program brands, 3) increased repeat-purchases, 4) increased buying frequency, 5) higher probability to be entirely loyal to the program brands, 6) higher probability to switch only among the program brands, not to non-program brands. Already a lot of research is done regarding the effect of LPs on behavioural loyalty. Using the division of Sharp and Sharp (1997), a lot of research is done about the effectiveness of LPs in terms of increased buying frequency of customers (Liu 2007; Sharp and Sharp 1997; Waarden and Benavent 2007). This research is different, since it focuses on type 2: ‘increased purchases/(start) buying at the program brand’. Sharp and Sharp (1997) mention in this effect both the event that existing customers increase their spending and the event that (new) customers (start) buying at the program brand. Since both elements (new and existing customers) are interesting for businesses, this type of behavioural loyalty is used during this research to assess the effect of the MVLP.

2.3 (MULTI-VENDOR) LOYALTY PROGRAMS

A way to enhance customer loyalty is the use of loyalty programs (LPs) (Bijmolt, Dorotic and Verhoef 2010). A LP is defined as: a program that tries to create customer loyalty by offering an incentive/benefit to current customers (Sharp and Sharp 1997). Looking at the four different phases of loyalty (as described in section 2.1), such a program mainly focuses on the creation of loyalty in stage three. In the third stage, a customer creates a commitment towards a product/service. This commitment can be enhanced or created through the opportunity for customers to get a reward for their purchases. In other words, there is a benefit for a consumer when he/she becomes committed to the product/service. This type of loyalty is tried to be created with the aim to create action loyalty, i.e. behavioural loyalty.

(12)

12

frequency reward programs, there are multiple designs. Some businesses introduce their own loyalty program, known as stand-alone programs (SAPs) (Rese, Hundertmark, Schimmelpfennig and Schons 2013). This type of program offers the customers the opportunity to be rewarded for their behaviour within that one specific business, e.g. a customer get rewarded through the LP of supermarket X for their behaviour at supermarket X. Another possibility is to join an existing LP. These LPs are known as multi-vendor loyalty programs (MVLPs). Based on the definition of Blattberg et al. (2008) and Dorotic et al. (2010), a MVLP is defined as: a type of LP which includes a coalition of not competing firms (i.e. vendors), managed by an independent specialized operator. It is important to make a distinction between both programs, because MVLPs do have some extra features in comparison to SAPs. From a customer perspective, MVLP makes it possible to collect points at multiple companies, allowing them to collect points quicker. It offers also the possibility to redeem the points at multiple partners (Dorotic et al. 2010. From a firm perspective, a MVLP gives a company the opportunity to benefit from the network within the program. This means that by participation not only the behaviour of current customers will be influenced, but a new partner can also acquire new customers because of the customer lock-in when they are a member of the MVLP (Rese et al. 2013).

Already some research is done regarding the effectiveness of MVLPs (Dorotic et al. 2010; Rese et al. 2013), but never before a distinction is made between different kind of collaborations between the MVLP and its partner(s). This research contributes to the existing literature by adding two different types of collaborations between the parties: ‘exchange partners’ and ‘exchange and save partners’. The first group, exchange partners, are companies within the MVLP at which it is only possible to redeem the collected points. The second group, exchange and save partners, are companies at which customers can collect and redeem their points. At some MVLP it might be possible to become only a save partner, but this type of collaboration is out of scope for this research.

(13)

13

Yi and Jeon 2003). A determinant of a higher valued program is the structure of the reward, more specific the reward frequency (McCall and Voorhees 2010). This means that when rewards are made available more often for the customer, it directly influences their participation in a LP (Smith and Sparks 2007). When a business decides become both an exchange and save partner, rewards become more accessible. This results in the possibility for the customer to collect points more quickly (reward frequency increases) and to become more involved. Since different effects are found for the level of customer involvement (Yi and Jeon 2003), it is interesting to make the distinction between the two collaboration types.

2.4 THE EXPECTED INFLUENCE OF A MVLP ON LOYALTY

Looking at behavioural loyalty, it consists of two elements: the current customers buying more (spend a higher amount of money per month) and the new customers start buying at the company. It is expected that a MVLP positively influence both elements, but a different level in the effect (not in the sign of the effect) on behavioural loyalty between both collaboration types. The next figure visualizes the expected relationship. The expected relationships are explained further in the next two paragraphs, resulting in the belonging hypotheses.

Figure 2.1 The expected main effect

(14)

14

through becoming a partner of a MVLP the company can benefit from the level of involvement which is already created within the program. Therefore, the first part of the first hypothesis is stated as follows:

Hypothesis 1A: Becoming an exchange partner of a MVLP has a positive influence on the

monthly spending of the partner’s customers

Another factor which is expected to influence the spending behaviour is the point pressure mechanism (Blattberg et al. 2008). This point pressure mechanism explains the change in customer behaviour when the customer nears the amount of points required for a reward (Kivetz, Urminsky and Zheng 2006). This mechanism assumes that customers are ‘goal-gradient’, meaning that people do accelerate their spending when they approach their goal (Hull 1932), i.e. the amount of points needed. It is assumed that this effect becomes even stronger due to the feature of MVLPs to collect points more quickly (Dorotic et al. 2010). Because it becomes possible for customers to collect points when a company becomes both an exchange and save partner, it is expected that this point pressure mechanism positively influences the spending behaviour of its customers. In addition, customers can collect points more quickly when a new save partner is added which might even strengthen this influence. Since it remains possible to hand in points for a discount during this type of collaboration with a MVLP, it is expected that this also positively influences customer spending behaviour (explained in previous paragraph). Based on this reasoning, the second part of the first hypothesis is stated as follows:

Hypothesis 1B: Becoming an exchange and save partner of a MVLP has a stronger positive

influence (compared to only an exchange partner) on the monthly spending of the partner’s customers

(15)

15

expected that the participation in a MVLP is positively related to the number of customers. Into more specific, it is expected that becoming an exchange partner does have a positive influence on the number of customers due to the fact that consumers might choose the MVLP partner’s to exchange their points for discount. This is expected since promotion does positively influence store preference (Smith and Sinha 2000). But this effect is even stronger when a company does become both an exchange and save partner. Besides the effect described in the previous sentences, the number of customers is expected to be also positively influenced since customers choose the MVLP partner’s to collect points because they are locked within the program (Hartmann and Viard 2007). This means, that customers choose specific companies since they can collect points at their transaction. Based on this reasoning, it is expected that becoming both an exchange and save partner has a stronger positive effect on the number of customers, than becoming only an exchange partner.

Since it is nowadays easy to buy products/services from any company, due to evolution of the internet, it becomes very convenient to switch to another provider (Rad, Ghorabi, Rafiee and Rad 2015). This means that customer are not bounded with a company anymore, resulting in the fact that every internet user is a potential buyer. Therefore, no distinction is made between customers who buy for the first time or for the nth time. This results in measuring behavioural loyalty regarding the number customer on a monthly basis. Therefore, the second hypothesis is stated as follows:

Hypothesis 2A: Becoming an exchange partner of a MVLP has a positive influence on the

number of the partner’s customers per month

Hypothesis 2B: Becoming an exchange and save partner of a MVLP has a stronger positive

influence (compared to only an exchange partner) on the number of the partner’s customers per month

2.5 REDEMPTION BEHAVIOUR

(16)

16

Blattberg et al. (2008) do mention three mechanisms by which customer value (i.e. customer spending and/or customer purchase frequency) can be increased: point pressure (section 2.2), personalized marketing (section 2.3) and rewarded behaviour (this section).

The rewarded behaviour mechanism is defined as: the increase in the purchase size and/or frequency due to the fact that a customer has received a reward (Blattberg et al. 2008). The change in the purchase behaviour is based on behavioural learning, which suggests that it is more likely for rewarded behaviour persist and holds (Blatterberg and Neslin 1990). Taylor and Neslin (2005) also found that it can also increase the commitment towards a business, which results in higher purchase rates. Based on the findings in the current literature (Kivetz and Simonson 2002; Lal and Bell 2003; Smith and Sparks 2009; Taylor and Neslin 2004), it is expected that the redemption of points does trigger the rewarded behaviour mechanism. This suggests than when a customer did redeem points at a specific point in time, their behaviour will be changed afterwards and remain changed. In short, the redemption of points has a positive effect on the relationship between a collaboration with a MVLP and the behavioural loyalty of its partner’s customers, arising from the moment of redemption and remain afterwards. Therefore it is expected that:

Hypothesis 3A: The redemption of points by a customer strengthens the effect of being a

MVLP partner on the partner’s customers monthly spending.

Hypothesis 3B: The redemption of points by a customer strengthens the effect of being a

MVLP partner on the number of customers.

(17)

17

value attached to the point by the customer). Based on this reasoning and the corresponding literature, it is suggested that:

Hypothesis 4A: The redemption of points by a customer strengthens the effect of being a

MVLP partner on the partner’s customers monthly spending more strongly when the (discount) value per point is increased.

Hypothesis 4B: The redemption of points by a customer strengthens the effect of being a

MVLP partner on the number of customers more strongly when the (discount) value per point is increased.

2.6 POINT COLLECTION

As mentioned in section 2.2, two different collaboration types are possible between the MVLP and its partner. Both types of collaboration make it possible to redeem points at transactions (described in section 2.4). But when a partner becomes both an exchange and a save partner, it becomes possible to collect/save points at their transaction. The option to collect points can create customer lock-in (Hartmann and Viard 2007). This lock-in occurs, because consumers adjust their store choice to the involved partners. They adjust their store preference, due to the option to collect points. This adjustment causes consumers to switch to another store/brand, which is even more easy in an online setting where it is very convenient to switch between suppliers (Red et al. 2015). Therefore, it is expected that the lock-in positively influences the number of customers.

This lock-in is even strengthened by the psychological phenomon mentioned before: the point-pressure mechanism (Taylor and Neslin 2004). The point-point-pressure mechanism (explained in detail in section 2.4) is expected to positively influence the number of customers, since the partner of the MVLP will attract members of the program due to the option to collect points at their business. Besides, this effect will remain over time since customers can collect points more quickly (due to a new option to collect points), which will even strengthen the point-pressure mechanism. Based on this findings, the fifth hypothesis is stated as follows:

Hypothesis 5: The collection of points by a customer does have a lasting positive moderating

(18)

18

2.7 CONCEPTUAL MODEL

This research investigates which effects a collaboration with a MVLP has on two components of the customer’s behavioural loyalty, namely: the number of customers and the monthly spending per customer. It is expected that this relationship is influenced by the redemption and collection of points by the customer. In figure 2.1 the expected relationship is summarized in a conceptual model.

Figure 2.2 Conceptual model

(19)

19

3. METHODOLOGY

In this section, the methodology of this research is discussed. First the data collection is discussed, followed by a description of the structure and storage of the dataset in section 3.2. Section 3.3 provides a description of the variables. In section 3.4 the model which is investigated during this research is discussed, followed by the representativeness of the model in section 3.5.

3.1 DATA COLLECTION

As discussed in chapter 2, this research investigates the effect of a collaboration with a MVLP on the behavioural loyalty of its partner’s customers. Into more detail, it focuses on the effect of MVLP XXX on the behavioural loyalty of the customers of ***. In order to investigate the proposed research questions, two different databases are used and combined with each other. The first database which is used to collect the data from, is the database of XXX. This database consists of cardholder data, like: name, address, etc. The second database which is used, is the database of ***. This database consists of information about their customers, like name, age, gender, etc. But also the amount of money their customers spend in Euro`s.

By combining the databases, an internal data source (i.e. XXX data) is combined with an external data source (i.e. *** data). This combined dataset provides a more powerful profile of the customer (Verhoef, Kooge and Walk 2016). It is possible to combine these two data sources since both consist of structured data, which can be linked through customer identity. This means that the dataset which is used for analysis consists of customer level data. Figure 3.1 shows how the two datasets are linked with each other.

Figure 3.1 Illustration of how the datasets are linked

(20)

20

date of birth, gender and postal code are equal in both sets. When a match is identified, the customer data from the XXX database is linked to the customer data from the *** database.

3.2 DATA STRUCTURE AND STORAGE

As described in the previous section, the data is collected through two different databases. Both databases are structured on the customer’s transaction, meaning that one customer can have multiple rows (if an unique customer did bought multiple times). In order to investigate the changes in the customer behaviour, the design is changed to panel data structure. Panel data is a data set that follows individuals over time, providing observations for each customer for every period in time (Hsiao, 2003). The level of analysis is set to monthly data. Since the data provides observations from 01-09-2013 till 30-06-2016, each customer has 34 rows. Where each row represents their behaviour in that specific month. The next two sub paragraphs specify the different variables which will be included in the analysis.

During this research the data is stored and tested, using the software Stata MP. Stata is a statistical software for data analysis. The version MP is the fastest version of Stata, which allows the user to analyse the data much quicker also with larger datasets. This software is chose for two reasons: 1) the software has a high capacity which is needed when working with large datasets and 2) the software is able to work with different data structures, i.e. it is able to recognize the panel data setting. It is important that the software does recognize this type of data structure, since there must be corrected for the multiple observations per customer. This correction deals with the treatment of heterogeneity in units (Leeflang et al. 2015). In other words, Stata must recognize that an unique customer does have multiple observations instead of counting all observations as different customers. This is done through the 𝑋𝑇𝑖𝑡 in Stata,

where 𝑖 does represent the unique customer (identified through the unique *** customer number) and 𝑡 the observation month.

3.3 MEASUREMENT OF THE CONSTRUCTS

(21)

21

3.3.1 VARIABLES

In order to measure the expected relationships, multiple variables are needed. Table 3.1 does provide all variables used in this research for testing the hypotheses. Some variables do require some additional information. When this is the case, this information is given afterwards. Those variables are indicated with a footnote where the number of the footnote does correspond with the number of the additional information.

Construct Variable name Variable type Measurement Reference level Buying – Indicating if a

customer made a purchase in the observation month.

Buying Binary (0 – 1)

1: for all observations in which a purchase is made

0: for all other observations

The zero’s in this variable

Monthly 𝐬𝐩𝐞𝐧𝐝𝐢𝐧𝐠(𝟏)

-The amount of Euro’s spend per month in all 34 months by a customer.

Monthly_ spending

Interval (0 - ∞)

> 0: transaction, value is the amount (€) spend in this specific month

0: no transaction

N.A.

Collaboration 𝐭𝐲𝐩𝐞(𝟐) –

Variables indicating which type of collaboration is applicable for that specific point of time.

No_collaboration Binary (0 – 1)

1: for all observations from September 2013 till July 2014 (incl. July 2014)

Yes, so left out

Exch Binary (0 – 1)

1: for all observations from August 2014 till April 2016 (incl. April 2016)

No

Exch_save Binary (0 – 1)

1: for all observations in May and June 2016

No

Point 𝐫𝐞𝐝𝐞𝐦𝐩𝐭𝐢𝐨𝐧(𝟑) -

An indicator whether a customer did redeem XXX (i.e. points) at their transaction

Redemption Binary (0 – 1)

1: for all observations when points were redeemed

0: for all observations when no points were redeemed

Increased

𝐩𝐨𝐢𝐧𝐭 𝐯𝐚𝐥𝐮𝐞(𝟒) - Indicator whether the redeemed points were increased in value

Redemp_ increased

Binary (0 – 1)

1: for all observations were redeemed points were increased in value

Points collected – An indicator whether a customer

Collected Binary (0 – 1)

1: for all observations were points were collected

(22)

22

did collect XXX (i.e. points) at their transaction Exchange partner * Points redeemed – Interaction effect Exch_Redemp Binary (0 – 1)

1: for all observations during the exchange period, starting from the moment when points were redeemed till the end of this period 0: for all other observations

The zero’s in this variable Exchange partner * Points redeemed * increased 𝐯𝐚𝐥𝐮𝐞(𝟓) – Interaction effect Exch_Redemp_ increased Binary (0 – 1)

1: for all observations during the exchange period, starting from the moment when points with increased value were redeemed till the end of this period

Exchange and save partner * Points redeemed – Interaction effect

Exch_save_ Redemp

Binary (0 – 1)

1: for all observations during the exchange and save period, starting from the moment when points were redeemed till the end of this period 0: for all other observations

Exchange and save partner * Points collected – Interaction effect

Exch_save_ Collected

Binary (0 – 1)

1: for all observations during the exchange and save period, starting from the moment when points were collected till the end of this period 0: for all other observations

Table 3.1 Variables used to measure the construct

1. Number of customers per month. As can be seen in table 3.1 no variable is created to measure the number of customers per month. This is the case since this number will be obtained through the variable monthly spending. When the monthly spending is above zero, i.e. the customer bought something, this observation will be counted as a customer in that specific month.

(23)

23

Figure 3.2 Observational timeline, along with the different collaboration types

3. Redemption. Since a customer is not required to hand in a minimum amount of XXX, a cut-off value is determined. This is done since this research want to measure the effect of a discount which is noticed by the customer. Joglekar (1988) found that the minimum amount of discount to be effective is 1%. But when a customer hands in their XXX for discount, the endownment effect appears (Thaler, 1980). This means that the one who ‘sells’ their points, i.e. the customer who hands in their XXX, addresses a higher value to it than the ‘buyer’, i.e. *** who gives a determined discount. Therefore it is assumed that the customer will value/notice the discount more when he/she did decide to hand in the points, since they already had to overcome the endownment effect. Therefore the cut-off value is 0,5% of the total transaction. When the value of the discount lies beneath 0,5% of the total transaction, the value of the variable Redemption will turn to zero, i.e. no redemption.

4. Increased point value. The increased value is identified through the combination of the two datasets. When the discount given in the *** data was not equal to the amount of XXX divided by 100 (normal value of one XXX is €0,01), it indicated an increased value.

5. Exchange partner * points redeemed * increased value. This interaction effect is not created for the partner type exchange and save, since no points were increased in value during this period.

3.3.2 CONTROL VARIABLES

(24)

24 Construct Variable name Variable type Measurement Reference level (only applicable when variable type is binary) Gender – A variable indicating the gender of the customer

C_Male Binary (0 – 1)

1: for male customers 0: for all other customers

Yes, so left out C_Female Binary

(0 – 1)

1: for female customers 0: for all other customers

No C_Gender_

Unknown

Binary (0 – 1)

1: for those customers whose gender is unknown

0: for all other customers

No

Age – Variables created to indicate to which age group the customer belongs. Note: this is exactly the same over the three years of observation

C_Age_1 Binary (0 – 1)

1: for customers who are younger than 30 years old

Yes, so left out

1: for customers who in the range of 30 years old till 39

No

1: for customers who are 70 years old or older

No

1: for customers whose age is unknown

No

Month – Variables are created indicating in which month the value is observed

C_Jan Binary (0 – 1)

1: for all observations in the month January

Yes, so left out

C_Feb Binary (0 – 1)

1: for all observations in the month February

(25)

25

C_Mar Binary (0 – 1)

1: for all observations in the month March

No

The same variables are created for the remaining months: April, May, June, July, august, September, October, November, December

Membership XXX – Variables indicating if a customer was an XXX member in a specific month C_Member Binary (0 – 1)

1: for all customers who were an XXX member in the month (along with the year) of observation 0: for all customers who are not an XXX member in the month (along with the year) of observation

Note: since it is only of interest if a customer is a member after *** became a partner, none of the customer will have a value of one before August 2014.

The zero’s in this variable (i.e. the non-members)

Mail Deal (𝐌𝐃)(𝟏)

– Variables indicating who did receive an email, who were allocated to the control group and who did not receive an email

C_MD_CG Binary (0 – 1)

1: for all observations were a customer was allocated to the control group

Yes, so left out

C_MD_AG Binary (0 – 1)

1: for all observations were a customer receive the email containing an MD

No

C_MD_NG Binary (0 – 1)

1: for all observations who did not received an MD and were not allocated to the control group 0: for all other observations

No

Exlusive 𝐃𝐞𝐚𝐥𝐬(𝟏)_–

Variables indicating who did receive an email, who were allocated to the control group and who did not receive an email

C_EXCL_CG Binary (0 – 1)

1: for all observations when a customer was allocated to the control group

Yes, so left out

C_EXCL_AG Binary (0 – 1)

1: for all observations when a customer receive the email containing an exclusive deal 0: for all other observations

No

C_EXCL_NG Binary (0 – 1)

1: for all observations who did not receive an exclusive deal and were not allocated to the control group 0: for all other observations

(26)

26 Email 𝐦𝐢𝐬𝐬𝐢𝐧𝐠(𝟐)_–

for not all month, an observation regarding email was available

C_EM Binary (0 – 1)

1: for all observations in the period when email observations were not available

Table 3.2 Control variables used during this research

1. Marketing XXX. This research controls for the influence on behavioural loyalty due to marketing activities by XXX. In order to control this influence of marketing on customer behaviour as precise as possible, but also to see whether the XXX marketing activities does have any influence (for further advice), different variables are created. There is a distinction made between the type of emails which were send to the XXX customers: 1) Mail Deal (MD) – These emails are a combination of offers from multiple partners, 2) Exclusive (EXCL) – These emails do only contain information about ***. The difference between the two emails is the level of attention for ***. This distinction is made since a consumer’s attention is limit (Pieters and Wedel 2004), which can influence the effect of the email on the customer’s behaviour. Therefore, different effects are expected between both types of personalized marketing.

2. Email Missing. *** became a partner of the MVLP of XXX in August 2014, but the observation period starts in September 2015. Thus, for the period before August 2014, no observations of personalized marketing from XXX are available. Besides, XXX changed from software supplier in February 2015. This also resulted in missing observations for the period until February 2015. In order to capture the unobserved effect in these months, a control variable is created which gives a value of one to the month in which the observations of personalized marketing are missing.

3.4 MODEL SPECIFICATION

(27)

27

indicates that a Tobit analysis is advised. Another aspect why this analysis is suitable, is the fact that the size of the spending behaviour will not go below the value of 0 (Fox, Montgomery and Lodish 2004). Due to the censoring in the data, the estimates of the model will be influenced when using ordinary least squares, i.e. the estimates will be biased (Yoshimoto 2008). The Tobit model overcomes this problem. The model do allow for dependent variables which are continuous and censored (Leeflang, Wieringa, Bijmolt and Pauwels 2015). Using this type of modelling, only the observation of those who purchased are used for estimation, leaving all observed zero’s out.

Multiple types of this analysis can be specified, two examples are: Type-1 Tobit and Type-2 Tobit models (Leeflang et al. 2015). The difference between both types lies within the prediction of the dependent variable. The Type-1 Tobit model uses the same explanatory variables for both explaining whether a customer makes a purchase and the size of its purchase. In the Type-2 Tobit model, two different sets of explanatory variables are used. In this research the Type-2 Tobit model will be used. This type seems most suitable since it is expected, based on the literature, that there are different variables influencing the decision where to purchase in comparison to the size of the purchase. A customer might choose to buy at a certain supplier due to the fact that they are rewarded for this choice, but this does not mean that a customer will spend more. Looking at the airline industry, this reasoning holds since the product of interest, i.e. airline tickets, are generally spoken well considered purchases instead of impulse buying items (Rook 1987). Therefore, this research assumes that customer are less/not seduced to buy additional products or other items on impulse (Hoch and Loewenstein 1991), influencing the size of the purchase. Resulting in an increased likelihood that the customer is well informed before a purchase is made and becomes less likely to buy what he/she is intended to buy (Jones, Reynolds, Weun and Beatty 2003). Based on this reasoning, it is expected that the number of customers is more influential that the size of the purchase. Therefore, a Type-2 Tobit model (here after Tobit model) seems most justified and therefore will be used to investigate the effect of the MVLP on customer behavioural loyalty.

3.4.2 TOBIT TYPE II/HECKMAN CORRECTION

(28)

28

panel data structure (Puhani 2000). The Heckman 2-step correction contains three steps: 1) estimating the probability of buying using a probit model, 2) estimating the inverse mills ratio of the probit model and 3) truncated regression with controlling for the truncation through adding the inverse mills ratio in the model (Hamelton and Nickerson 2001).

Step 1. Probit model.

A probit model is a binary choice model (Leeflang et al. 2015). This analysis does provide information about whether a customer bought something or not, e.g. the number of customers per month. This is identical to the first step in the Tobit model: accounting for whether the dependent variable turns zero or one (Leeflang et al. 2015). Since the data used for this model has a panel data setting, the XTPROBIT analysis in Stata is suitable. The next equation does represent the factors influencing a customer to buy or not.

𝐵𝑢𝑦𝑖𝑛𝑔_𝑖𝑡∗ (1) = 𝐸𝑥𝑐ℎ𝑡𝛽(1𝐴)+ 𝐸𝑥𝑐ℎ_𝑆𝑎𝑣𝑒𝑡𝛽(1𝐵)+ 𝐸𝑥𝑐ℎ_𝑅𝑒𝑑𝑒𝑚𝑝𝑖𝑡𝛽(1𝐶) + 𝐸𝑥𝑐ℎ_𝑅𝑒𝑑𝑒𝑚𝑝_𝑖𝑛𝑐𝑟𝑒𝑎𝑠𝑒𝑑𝑖𝑡𝛽(1𝐷) + 𝐸𝑥𝑐ℎ_𝑆𝑎𝑣𝑒_𝑅𝑒𝑑𝑒𝑚𝑝𝑖𝑡𝛽(1𝐸) + 𝐸𝑥𝑐ℎ_𝑆𝑎𝑣𝑒_𝐶𝑜𝑙𝑙𝑒𝑐𝑡𝑖𝑜𝑛_𝑖𝑡𝛽(1𝐹)_{+ 𝐶_𝐺𝑒𝑛𝑑𝑒𝑟} 𝑖𝛽(1𝐺)+ 𝐶_𝐴𝑔𝑒𝑖𝛽(1𝐻) + 𝐶_𝑀𝑜𝑛𝑡ℎ_𝑖𝛽(1𝐼)_{+ 𝐶_𝑀𝐷} 𝑖𝑡𝛽(1𝐽)+ 𝐶_𝑀𝐷_𝑁𝐺𝑖𝑡𝛽(1𝐾)+ 𝐶_𝐸𝑥𝑐𝑙𝑖𝑡𝛽(1𝐿) + 𝐶_𝐸𝑥𝑐𝑙_𝑁𝐺𝑖𝑡𝛽(1𝑀) + 𝐶_𝐸𝑀𝑡𝛽𝑡 (1𝑁) + 𝐶_𝑀𝑒𝑚𝑏𝑒𝑟𝑖𝑡𝛽(1𝑂)+ 𝛾(1)+ 𝜀(1)

Note: The direct effect of those variables included in the moderation effect are left out, since they will have the same value as the moderation effect due to the fact that they are dummy variables.

Since the outcome of the probit model is hard to interpret (Risselada 2015), the marginal effects per variable must be obtained. A marginal effect is the effect of a change in a variable on the probability. Since all explanatory variables are dummy variables (i.e. binary variables), the marginal probability effect is obtained through the next equation:

Φ (𝑋_1𝑖𝑡𝛽) − Φ (𝑋_0𝑖𝑡𝛽) Where,

𝛽 represents the estimated coefficient of variable 𝑋𝑖𝑡

Φ (𝑥1𝑖𝑡𝛽) represents the probability when variable 𝑋𝑖𝑡 = 1 (i.e. dummy equals one),

holding all other variables constant

(29)

29

Thus, a marginal effect represents the change in the probability when an explanatory variable changes in value. In this case: what is the effect of the explanatory dummy variable changing from the value of zero to the value of one, on the probability of buying.

Step 2. Inverse mills ratio.

In order to account for the censoring in the latent variables in the size of the spending, the inverse Mills Ratio is calculated (Heien and Wessells 1990). This means that by including this ratio as an independent variable in the regression, it controls for the selection bias (i.e. the probability of buying or not). This ratio is obtained by dividing the probability density function (hereafter pdf) by the cumulative density function (hereafter cdf) (Puhani 2000). The values for both the pdf and cdf are obtained through Stata, using the following equation:

𝐿𝑎𝑚𝑏𝑑𝑎 = 𝜙 (𝑋𝑝𝑟𝑜𝑏𝑖𝑡) Φ (𝑋_{𝑝𝑟𝑜𝑏𝑖𝑡)} Where,

𝐿𝑎𝑚𝑏𝑑𝑎 represents the outcome of this equation is saved as a new variable

𝑋_{𝑝𝑟𝑜𝑏𝑖𝑡} represent all variables used to predict the event of buying (see section 4.2.1) 𝜙 represents the pdf of the estimated probit model

Φ represents the cdf of the estimated probit model

Step 3. Truncated regression.

A truncated regression is characterized by a regression in which not all observations are used (Leeflang et al. 2015). The observations which are not used, are not observed. This type of analysis is used since the transformation to panel data created rows for those observations in which a purchase is made and not made. By using a truncated regression the observations in which no purchase is made, i.e. those observation which are observed, are not used. In order to account for the truncation in the data, three settings are changed in comparison to standard regression: 1) only those observations for which monthly spending ≠ 0 are used, 2) lambda is added as explanatory variable to account for the truncation and 3) maximum likelihood method is used (Park and Simar 2008). Since the data has a panel data setting, the XTREG analysis with the condition ‘if monthly spending > 0’ in Stata is suitable. The next equation explains the size of a customers’ spending.

𝑆𝑝𝑒𝑛𝑑𝑖𝑛𝑔_𝑖𝑡∗ (2) = 𝐸𝑥𝑐ℎ_𝑡𝛽(2𝐴)_{+ 𝐸𝑥𝑐ℎ_𝑆𝑎𝑣𝑒}

𝑡𝛽(2𝐵)+ 𝐸𝑥𝑐ℎ_𝑅𝑒𝑑𝑒𝑚𝑝𝑖𝑡𝛽(2𝐶)+

(30)

30 𝐶_𝐴𝑔𝑒_𝑖𝛽(2𝐺)_{+ 𝐶_𝑀𝑜𝑛𝑡ℎ}

𝑖𝛽(2𝐻)+ 𝐶_𝑀𝐷𝑖𝑡𝛽(2𝐼)+ 𝐶_𝑀𝐷_𝑁𝐺𝑖𝑡𝛽(2𝐽)+ 𝐶_𝐸𝑥𝑐𝑙𝑖𝑡𝛽(2𝐾)+

𝐶_𝐸𝑥𝑐𝑙_𝑁𝐺_𝑖𝑡𝛽(2𝐿) + 𝐶_𝐸𝑀_𝑡𝛽_𝑡(2𝑀)+ 𝐶_𝑀𝑒𝑚𝑏𝑒𝑟_𝑖𝑡𝛽(2𝑁)+ 𝐿𝑎𝑚𝑏𝑑𝑎_𝑖𝑡𝛽(2𝑂)+ 𝛾(2)+ 𝜀(2) .

Note: The direct effect of those variables included in the moderation effect are left out, since they will have the same value as the moderation effect due to the fact that they are dummy variables.

1. Complete model.

The different models described in the previous section can be combined into one model, following the Tobit notation.

𝑌_𝑖𝑡(2) = 𝑌_𝑖𝑡∗(2) if 𝑌_𝑖𝑡∗(1) > 0 = 0 if 𝑌_𝑖𝑡∗(1) ≤ 0 where, 𝐵𝑢𝑦𝑖𝑛𝑔_𝑖𝑡∗ (1)= 𝐸𝑥𝑐ℎ_𝑡𝛽(1𝐴)_{+ 𝐸𝑥𝑐ℎ_𝑆𝑎𝑣𝑒} 𝑡𝛽(1𝐵)+ 𝐸𝑥𝑐ℎ_𝑅𝑒𝑑𝑒𝑚𝑝𝑖𝑡𝛽(1𝐶)+ 𝐸𝑥𝑐ℎ_𝑅𝑒𝑑𝑒𝑚𝑝_𝑖𝑛𝑐𝑟𝑒𝑎𝑠𝑒𝑑_𝑖𝑡𝛽(1𝐷)+ 𝐸𝑥𝑐ℎ_𝑆𝑎𝑣𝑒_𝑅𝑒𝑑𝑒𝑚𝑝_𝑖𝑡𝛽(1𝐸)+ 𝐸𝑥𝑐ℎ_𝑆𝑎𝑣𝑒_𝐶𝑜𝑙𝑙𝑒𝑐𝑡𝑖𝑜𝑛_𝑖𝑡𝛽(1𝐹)+ 𝐶_𝐺𝑒𝑛𝑑𝑒𝑟_𝑖𝛽(1𝐺)+ 𝐶_𝐴𝑔𝑒_𝑖𝛽(1𝐻) + 𝐶_𝑀𝑜𝑛𝑡ℎ_𝑖𝛽(1𝐼)+ 𝐶_𝑀𝐷_𝑖𝑡𝛽(1𝐽)_{+ 𝐶_𝑀𝐷_𝑁𝐺} 𝑖𝑡𝛽(1𝐾)+ 𝐶_𝐸𝑥𝑐𝑙𝑖𝑡𝛽(1𝐿)+ 𝐶_𝐸𝑥𝑐𝑙_𝑁𝐺𝑖𝑡𝛽(1𝑀) + 𝐶_𝐸𝑀𝑡𝛽𝑡 (1𝑁) + 𝐶_𝑀𝑒𝑚𝑏𝑒𝑟𝑖𝑡𝛽(1𝑂)+ 𝛾(1)+ 𝜀(1) and 𝑆𝑝𝑒𝑛𝑑𝑖𝑛𝑔_𝑖𝑡∗ (2) = 𝐸𝑥𝑐ℎ_𝑡𝛽(2𝐴)_{+ 𝐸𝑥𝑐ℎ_𝑆𝑎𝑣𝑒} 𝑡𝛽(2𝐵)+ 𝐸𝑥𝑐ℎ_𝑅𝑒𝑑𝑒𝑚𝑝𝑖𝑡𝛽(2𝐶)+ 𝐸𝑥𝑐ℎ_𝑅𝑒𝑑𝑒𝑚𝑝_𝑖𝑛𝑐𝑟𝑒𝑎𝑠𝑒𝑑_𝑖𝑡𝛽(2𝐷)+ 𝐸𝑥𝑐ℎ_𝑆𝑎𝑣𝑒_𝑅𝑒𝑑𝑒𝑚𝑝_𝑖𝑡𝛽(2𝐸)+ 𝐶_𝐺𝑒𝑛𝑑𝑒𝑟_𝑖𝛽(2𝐹)+ 𝐶_𝐴𝑔𝑒𝑖𝛽(2𝐺) + 𝐶_𝑀𝑜𝑛𝑡ℎ𝑖𝛽(2𝐻)+ 𝐶_𝑀𝐷𝑖𝑡𝛽(2𝐼)+ 𝐶_𝑀𝐷_𝑁𝐺𝑖𝑡𝛽(2𝐽)+ 𝐶_𝐸𝑥𝑐𝑙𝑖𝑡𝛽(2𝐾)+ 𝐶_𝐸𝑥𝑐𝑙_𝑁𝐺_𝑖𝑡𝛽(2𝐿)_{+ 𝐶_𝐸𝑀} 𝑡𝛽𝑡 (2𝑀) + 𝐶_𝑀𝑒𝑚𝑏𝑒𝑟_𝑖𝑡𝛽(2𝑁)_{+ 𝛾}(2)_{+ 𝜀}(2)_. 3.5 REPRESENTATIVENESS

A model is a representation of reality, with the aim to comprehend this reality (Wieringa 2015). It should be kept in mind that it is very hard, or it may be even impossible to capture reality in a model. Important aspects which should be considered when interpreting the estimated model are the expected influence of biases and missing values.

3.5.1 BIASES

(31)

31

when the findings presented in this report are used for general statements about the effect of XXX on the behavioural loyalty of its partner’s customers, a sample selection bias may arise. This is expected since the customers of ***, along with their behaviour, do not represent the customers of a partner in for instance fast moving consumer goods. In other words, the customer of *** who is a member of XXX does not represent all members of the MVLP. It is important to pay attention to this bias since other research showed that effects of a LP are more visible in airline sectors (Lal and Bell 2003). Therefore, the sample selection bias should be kept in mind when taking the findings of this research into general perspective. In order to keep the influence of this bias limited, a variable is added which controls for XXX membership (see section 3.2.2).

Another bias which might be present in the data encompasses the fact that the period of the observation time regarding the collaboration type ‘exchange and save partner’ is very limited in comparison to the period when XXX was not involved within the business and the period when *** was only an ‘exchange partner’. Therefore one should keep in mind that when no effects are found, effect might only become observable overtime. But also, effects might become stronger over time. To control for this bias as much as possible, a control variable is added which captures the different industry effects per month (see section 3.2.2).

3.5.2 MISSING VALUES

Although the data for this research is provided by real databases, some values are missing. These missing values are present in variables regarding the customer’s characteristics, like: gender, age, mail. For this research ‘mail’ is not directly important, but since the two databases are linked with each other based on the combination of these three, missing links (between the two database) might appeared. This problem is tried to be solved by including customers who were not identified as XXX member, but who made a transaction with XXX (i.e. redeemed/collected points). For these customers the month in which the first transaction is made with XXX is taken as the start month of their membership. The missing values in the other two variables (gender and age), for which will be controlled in this research, are grouped together into a variable which encompasses the missing values (see section 3.2.2).

4. RESULTS

(32)

32

4.1 DATA EXPLORATION

Before the model is estimated, a description of the data which is used for the analysis is given. The dataset consists of 290.156 unique customers of ***, observed during the period September 2013 till June 2016. Two different characteristics are obtained through the databases, namely: gender and age. Table 4.1 describes the profile of the sample.

Characteristic Levels Percentage

Gender Male 50.6 %

Female 33.8 %

Unknown 15.6 %

Age Till 30 years old 15.8 %

30 – 39 years old 13.3 % 40 – 49 years old 17.8 % 50 – 59 years old 19.6 % 60 – 69 years old 10.6 % + 70 years old 3.9 % Unknown 9.0 %

Table 4.1 Data profile

As can be seen in table 4.1, the sample consists of more males than females and most customers are grouped in the age category: 50 – 59 years old. The differences do not cause any statistical problems since this research controls for both customer characteristics (see section 3.2.2). Among those customers, ### customers are identified as XXX member. Among the members, ### customers did redeem XXX at their transaction and ### customers collected XXX at their transaction. On average, customers bought ### times during the observation time. The distribution of the overall sales in this period is shows in figure 4.1.

(33)

33

As can be seen in figure 4.1, the sales of *** varies per month. In every year, a peak is observed in the month --. This fluctuation does suggest that there is a factor which influences the higher sales in this period in time. Based on the observed trend, it might be that customers of *** do book their summer vacation in the month --, causing a peak in the sales.

4.2 MODEL FREE EVIDENCE

Before the research model is estimated, the construct of interest of this research is explored. Exploring the construct does give already a feeling of the structures in the data, but it also gives a directive for the final model estimates (Keim 2001). This can serve as a reference whether the final model estimation is done properly. The construct is explored through comparing customers who do use the MVLP rewarding system (i.e. collect and/or redeem points) with the customers who do not, regarding the frequency and size of purchasing activities (i.e. behavioural loyalty).

The first insights is generated with respect to difference between the customers who never redeemed XXX (group A) and the customers who ever did (group B). The difference in behavioural loyalty (i.e. the frequency and size of transaction(s)) is measured with an independent-samples t-test. Frequency is measured as the sum of the transactions during the observation period in numbers, whereas size is measured as the sum of all transactions in Euro’s. Table 4.2 does provide the output of the test statistic.

Redemption behaviour Mean SD t p

Frequency Group A: No ### ### -16.731 <.001

Group B: Yes ### ###

Size Group A: No ### ### 3.492 <.001

Table 4.2 Independent-samples t-test redemption behaviour

Both tests are highly significant (p=<.001), which indicates that there is a significant difference between the two groups regarding behavioural loyalty. The analysis regarding buying frequency indicates that customers who did redeem XXX once, are … to buy at ***. The second test reveals that this group also significantly spend … money. A simple calculation reveals that customer who did ever redeem XXX do deliver ### revenue in 34 months, namely €### (€###*###) for non-redeemers and €### (€###*###) for redeemers.

(34)

34

collection behaviour does only influences the buying frequency, one independensamples t-test is performed. Table 4.3 does provide the outcome of this t-test.

Collection behaviour Mean SD t p

Frequency Group A: No ### ### -18.434 <.001

Table 4.3 Independent-samples t-test collection behaviour

The test results show that there is a significant (t=-18.434, p=<.001) difference between the number of purchases for group A (M=###, SD=###) and group B (M=###, SD=###). Based on this test, it is suggested that customers who do collect XXX are more likely to buy again/more often at ***. Looking at customer revenue, it is also expected that the customers who do collect XXX generate €### (€###*###) revenue and €### (€###*###) for customers who do not collect. This calculation suggests that customers who do collect generate €### more revenue in 34 months.

4.3 MODEL OUTPUT

In order to test the hypotheses, two models are estimated (as described in section 3.4): probit and regression model. The next two sub paragraphs do give insights into the output of both models respectively. This chapter ends with a calculation of the customer revenue, based on the estimated models.

4.3.1 PROBIT MODEL

To gain insights into the effect of the MVLP on the number of customers, a probit model is estimated using panel data settings (command: 𝑋𝑇_𝑅𝐸𝐺 in Stata). Before looking at the sign, size and significance of the variables, the fit of the model is discussed. This is done before any results are interpreted, since assessing the fit of the specified model is one of the most important step in modelling (Yuan 2005).

(35)

35

performed in order to get insights into the overall model significance. The LR test diagnoses whether the estimated model is better in predicting the event of customers buying than random selection (the null-model) based on the log-likelihood. The log-likelihood of the estimated model is gained through an iteration process, with the null-model as a starting point (see appendix A for the iteration process of the probit model). The next equation calculates the LR of the model:

−2(𝐿𝐿_{𝑛𝑢𝑙𝑙}− 𝐿𝐿_{𝑝𝑟𝑜𝑏𝑖𝑡}) = −2((−1577986.5) − (−1530648.2)) = 94676.6 Where,

𝐿𝐿_{𝑛𝑢𝑙𝑙} represents the log-likelihood of the null-model 𝐿𝐿_{𝑝𝑟𝑜𝑏𝑖𝑡} represents the log-likelihood of the estimated model

The test produced a statistic of 94676.6 for the LR of the probit model. The critical chi-square value, at a significance level of .01 with 31 degrees of freedom, is 59.7. Since the estimated value of LR exceeds the critical value, the model is significantly better than the null-model (i.e. random selection). In short, the predictor variables do have an effect on the probability of buying (i.e. the null hypothesis that the coefficients are equal to zero can be rejected).

Validation. Before interpreting the individual effects within the probit model it is important to check if the problem of multicollinearity is present. When this is the case the estimated values are unreliable (Wieringa 2015). In order to check for this problem, the VIF-values are obtained using a regression analysis since it is not possible to obtain these values in Stata from a probit analysis (Williams 2003). Although a regression analysis is not the analysis of interest, the values do give good insights in the multicollinearity between the used variables. When the VIF-value exceeds the limit of 7, the issue of multicollinearity is present (Wieringa 2015). Table 4.4 does give the values for the variables used in the probit analysis.

Variable VIF

Exchange collaboration period 2.702

Exchange and save collaboration period 2.768

Redemption during the exchange collaboration period 1.411 Redemption with increased point value during the exchange collaboration period 1.007 Redemption during the exchange and save collaboration period 1.578 Point collection during the exchange and save collaboration period 1.440

Female customers 1.107

Gender unknown customers 4.680

Customers aged between 30 – 39 years old 1.661

(36)

36

Customers age unknown 5.039

XXX members 1.820

Mail Deal with *** mail received 31.089

Mail Deal with *** mail not received 159.983

Exclusive *** mail received 1.471

Exclusive *** mail not received 6.409

Period with missing mail information 161.934

February 1.748 March 2.218 April 2.102 May 2.178 June 2.173 July 1.482 August 1.410 September 1.571 October 1.537 November 1.482 December 1.491

Table 4.4 VIF-values probit model

(37)

37

−2(𝐿𝐿𝑚𝑜𝑑𝑒𝑙 𝐴− 𝐿𝐿𝑚𝑜𝑑𝑒𝑙 𝐵) = −2((−1530648.2) − (−1530648.3)) = −0.2

Since the estimated LR value does not exceed the critical chi-square value of 10.8 (p <.01 and df 1), the model with the variable is not significantly better, i.e. model A. Since it is important to keep a model parsimonious, the model without the variable which indicates the period of missing mail observations will be used for further analysis.

Individual marginal effects. Since the overall model is assessed, the coefficients of the variables within the model can be validated. One should keep in mind that the estimates of this model are hard to interpret (Risselada 2016). But in order to do so, the marginal effect per variable are obtained. The marginal effects are represented in table 4.5. Note: in appendix D the output regarding the coefficients which optimize the maximum likelihood can be found which are used as input for obtaining the marginal effects.

Variable dF/dx

Collaboration type

No collaboration¹ N.A.

Exchange collaboration period¹ ###***

Exchange and save collaboration period¹ ###***

Redemption during the exchange collaboration period¹ ###*** Redemption with increased point value during the exchange collaboration

period¹ ###***

Redemption during the exchange and save collaboration period¹ ###*** Point collection during the exchange and save collaboration period¹ ###*** Gender

Male customers¹ N.A.

Female customers¹ ###***

Gender unknown customers¹ ###***

Age

Customers aged below 30 years old¹ N.A.

Customers aged between 30 – 39 years old¹ ###***

Customers age unknown ###***

XXX members¹ ###***

Mail

Mail Deal with *** mail control group¹ N.A.

Mail Deal with *** mail received¹ ###*

Mail Deal with *** mail not received¹ ###***

Exclusive *** mail control group¹ N.A.

Exclusive *** mail received¹ ###***

Exclusive *** mail not received¹ ###***

(38)

38 January¹ N.A. February¹ ###*** March¹ ###*** April¹ ###*** May¹ ###*** June¹ ###*** July¹ ###*** August¹ ###*** September¹ ###*** October¹ ###*** November¹ ###*** December¹ ###***

Table 4.5 Estimates of the probit model, note: constant is not given since this variable does not have a marginal effect. (¹) dF/dx is for discrete change of dummy variable from 0 to 1, *** p≤.01, ** p≤.05, * p≤.1, n.s. p>.1, NA: reference level

Main effect. As can be seen in table 4.5, almost all variables do significantly influence the probability of happening of the event ‘buying’. This means, that for those variables, the null hypothesis (coefficient equals 0) can be rejected. Looking at the variable representing the period in which *** became exchange partner of the MVLP in comparison to being no partner, the collaboration has a positive significant influence (p=<.001) on the buying probability. The probability of buying increases with ### percent point during this period of collaboration with the MVLP. Becoming an exchange and save partner does also significantly influence the buying probability (p=<.001). During this type of collaboration, the probability of buying increases with ### point percent. This means that the influence on the buying probability increase along with the involvement (from exchange partner to exchange and save partner) in the MVLP. Based on these results there is enough evidence to accept hypothesis 2A and 2B.

The effectiveness of a multi-vendor loyalty program on customer behavioural loyalty

The effectiveness of a multi-vendor loyalty program

on customer behavioural loyalty

The effectiveness of a multi-vendor loyalty program

on customer behavioural loyalty

TABLE OF CONTENTS

1. INTRODUCTION

2. THEORECTICAL FRAMEWORK

3. METHODOLOGY

4. RESULTS