• No results found

The Dutch Health insurance market The effect of perceived costs and benefits on churn intention and actual churn

N/A
N/A
Protected

Academic year: 2021

Share "The Dutch Health insurance market The effect of perceived costs and benefits on churn intention and actual churn"

Copied!
64
0
0

Bezig met laden.... (Bekijk nu de volledige tekst)

Hele tekst

(1)

The effect of perceived costs and benefits on

churn intention and actual churn

The Dutch Health insurance market

Miriam Hoekstra June 26, 2017

(2)

2

The effect of perceived costs and benefits on

churn intention and actual churn

The Dutch Health insurance market

Miriam Hoekstra University of Groningen Faculty of Economics and Business

(3)

3

MANAGEMENT SUMMARY

The accountability of the marketing department, a topic discussed by many and still up to date. The decreasing influence of the marketing department is noted by a lot of academics. However the contribution of the marketing department in improving customer relationships or gaining competitive advantage should not be underestimated. New insights and techniques need to be developed to regain the influence marketing deserves.

Churn management plays an important role in many service industries and is therefore highly important for firms. The Dutch health insurance industry copes with increased competition since the new legislation in 2006. In order to reduce or even prevent churn, more knowledge and insights in churn behavior need to be obtained. Gaining insights in churn probabilities is in this study done through the effect of churn intention. In most cases the intention to churn is driven by certain factors that drive behavior. Perceived benefits and costs are indicated as drivers for churn intention. An important step is to indicate if churn intention acts as a mediator between these benefits, costs and actual churn. The threat of churn should motivate insurers to be responsive to the customers’ needs, preferences and their heterogeneity. Therefore, a segmentation analysis is added to gain insights in the differences between the segments.

This study finds that the direct influence of churn intention on actual churn is present. When assessing the effects of the perceived benefits the most important driver for churn intention is linked to coverage. The fact that a customer receives a better coverage at another insurer indicates a high intention to churn. The perceived benefit premium also indicates to have a high influence on churn intention. When the premium of another insurer is low, a customer expresses a higher churn intention. Unfortunately due to limitations the effect of the perceived costs could not be analyzed. The mediating effect of churn intention is shown in seven out of ten independent variables, which is promising for firms. Finally the segments indicate differences in churn intention an actual churn. Segments that include young customers express a higher churn intention and a higher likelihood to churn. By knowing more about the benefits, levels of churn intention, actual churn and the differences per segments, firms have the opportunity to anticipate to customer behavior in an early stage, due to the insights in churn intention. Retention campaigns can be specifically target to those expressing high intentions, which will contribute in the departments’ accountability.

(4)

4

PREFACE

At first when I started the Pre-master I just wanted to gain more knowledge in the area of Marketing, take it a level further than my higher level applied education Commercial Economics. That year was a struggle, my health wasn’t optimal and I really needed to work hard to pass math and statistics. Then the Master period started. I chose the specification Marketing Intelligence, I already knew a lot about the management part and wanted to challenge myself. And so I did, a new ‘world’ opened up. A very interesting part of marketing, which raised my interest and made me eager to learn more about it. Now, the final step of my master Marketing Intelligence has arrived, the Master Thesis. After a first attempt in 2015, which I couldn’t go through with due to health issues, the moment is there. With proud I present my final thesis. It wasn’t always easy, but I’m proud of what I’ve accomplished and look forward to what is coming next in my career.

A Master Thesis always comes with some help, therefor I would like to thank some people. First of all I would like to thank my supervisor Edwin Kooge for his support, flexibility and useful feedback in the entire process. Next to that I would also like to thank my group member Johan Lems for his input, advice and brainstorm sessions. Also my dear friend Dilara Tuna deserves a big thanks for her support and feedback. Last but not least my boyfriend, his patience, motivation and understanding has been a great support during this period.

Miriam Hoekstra

(5)

5

TABLE OF CONTENT

1. INTRODUCTION ...7

1.1 Dutch health insurance market ...7

2. THEORETICAL FRAMEWORK ... 10

2.1 Churn ... 10

2.1.2 Churn intention ... 11

2.1.3 Churn in the Health Insurance industry ... 12

2.2 Perceived benefits ... 13 2.2.1 Price ... 13 2.2.2. Service quality ... 14 2.2.3. Package ... 15 2.3 Perceived costs ... 16 2.3.1. Procedural loss... 16 2.3.2 Financial loss ... 16 2.3.3. Relational loss ... 17

2.4 Segmentation based on customer characteristics ... 17

2.4.1 Demographic characteristics ... 18 2.4.2 Geographic ... 19 2.4.3. Channel usage ... 19 2.6 Conceptual model ... 21 3. METHODOLODY ... 22 3.1 Dataset ... 22

3.1.1 Restructuring the dataset ... 23

3.2 Variables ... 23

3.2.1 Transforming and recoding variables ... 23

3.3 Missing variables and oddities ... 25

3.3.1 Weighting ... 26

3.4 Churn analysis... 26

3.4.1 Factor analysis ... 27

3.4.2 Logistic regression analysis ... 27

3.4.2 Moderation analysis ... 27

3.4.3 Mediation analysis ... 28

3.7 Segmentation analysis ... 29

4. RESULTS ... 30

(6)

6

4.1.1 Demographic and geographic statistics ... 30

4.1.2 Channel usage ... 31

4.1.3 Churn intention and actual churn ... 31

4.1.4 Perceived costs ... 32

4.1.5 Perceived benefits ... 32

4.2 Factor analysis ... 33

4.3 Segmentation analysis ... 36

4.4 Regression analysis ... 39

4.4.1 Hypothesis 1a: Effect of churn intention on actual churn ... 39

4.4.2 Hypotheses 2a: Premium is most important perceived price benefit. ... 40

4.4.3 Hypotheses 2b: Premium is overall most important perceived benefit. ... 41

4.4.4 Hypothesis 3: Service quality positively affects churn intention ... 41

4.4.5. Hypothesis 4a: Coverage positively influences churn intention ... 42

4.4.6. Hypothesis 4b: Effect of supplementary insurance is stronger than coverage. ... 42

4.5 Moderation analysis ... 42

4.5.1 Hypothesis 9a t/m d: Demographics ... 42

4.5.2 Hypothesis 10a t/m d: Geographic characteristics ... 43

4.5.3. Hypothesis 12: Perceived benefits and churn intention is different per segment ... 44

4.5.4 Hypothesis 13: Effect of churn intention on churn is different per segment. ... 45

4.6 Mediation analysis ... 46

4.6.1 Hypothesis 1b: Churn intention acts a mediator... 46

4.6.2 Overview of hypothesis ... 47

5. DISCUSSION ... 48

5.1 General discussion ... 48

5.1.1 Churn and churn intention ... 48

5.1.2 Perceived benefits and costs ... 49

5.1.3 Segments ... 50

5.2 Limitations and implications future research ... 51

5.3 Managerial implications ... 51

Appendix A ... 56

Appendix B ... 58

(7)

7

1. INTRODUCTION

Accountability is up to today a highly discussed topic in the area of marketing. Like in all departments in a firm, expenditures should be accounted for. The influence of the marketing department within a firm is often underestimated and therefore is seen as irrelevant (Verhoef and Leeflang 2009). To gain more respect, firms need to experience the added value of marketing. Marketing contributes in improving customer relationships, gaining competitive advantage and enhancing the financial performance of a firm (Verhoef and Leeflang 2009). Nowadays more techniques are developed to increase accountability which gives marketing department the opportunity to regain their influence.

1.1 Dutch health insurance market

In 2006 a new legislation was introduced, called the Health Insurance Act (HIA). The main goal of the HIA was to stimulate the (price) competition between the different insurers in order to increase the quality, the accessibility and the affordability of healthcare. The Dutch health care system changed from a more paternalistic system to a system where customers gained more freedom to make their own decisions, which also led to more responsibilities (Nivel, 2016). One of the major changes since the introduction of the HIA was the increase in mobility. Approximately 18% of the customers churned to a different insurer, relative to 5% in the previous years. Besides that 10% of the customers made some changes in their insurance package within the same insurer. At the time this is seen as a positive impulse for the competitive pressure in this new liberalized market (Nza, 2006). Right after the reform, when the dust settled, churning rates stabilized up to 2010 and started increasing again up to last year.

(8)

8 firms. At this moment there are in total 25 Dutch health insurers, most of these are part of an overarching concern. In total there are nine overarching concerns. Of these nine concerns there are four dominant players with the largest market share; Achmea (30,38%), VGZ (24,07%), CZ (20,67%) and Menzis (13,35%). Each year these four dominant concerns lose a bit of market share to the smaller concerns (NZa, 2016). These numbers indicate that competition is increasing, which creates a better position for the smaller concerns.

All Dutch inhabitants are entitled to health care or compensation of health care, which is part of the standard package of the HIA. Health insurers need to offer at least one type of basic insurance with benefits specified by the law. All extras not covered by the basic package can be complemented with a supplementary insurance, for instance dental care or physiotherapy. The supplementary insurance is customized to the needs and preferences of each customer. Over the last years a decline in the supplementary insurances is noticed. In 2011 almost 90% of the customers had a supplementary insurance, while in 2015 this was 84%. Besides the basic and supplementary insurances, it is also possible to start an insurance contract as individual or as collective. A recent trend in the Dutch health insurance market are the so called budget insurance policies with mitigating conditions. In exchange for a lower premium customers agree with these mitigations for instance, limited choices from contracted healthcare providers or a lower compensation when non-contracted healthcare is used. In the last years an increase in these policies is noticed. In 2008 there was only one budget policy offered, while in 2015 there were in total 17 budget policies. The supply and demand of these policies is increasing since 2011, last year 1,23 million customers entered a budget contract (Nza, 2016).

(9)

9 churn behavior in the Dutch health insurance market. An analysis of the perceived costs and benefits indicate the drivers that influence behavior. Next to actual churn, the mediating effect of churn intention and the moderating effect of segmentation is analyzed. The moderating effect is added due to the fact that customers aren’t homogenous. The differences in for instance age, income and education are accounted for. By doing so, specific information about the differences between customers and their behavior is obtained. This valuable information is relevant for marketing purposes. It supports targeting and retaining the right customers, which increases the accountability of the marketing department. In marketing literature there is extensive research on costs and benefits that influence actual churn. However, the influence of churn intention seems to be either less relevant or underestimated in recent literature. This study contributes to existing literature by not only focusing on actual churn, but taking into account the effect of churn intention and the mediating role of this variable. By analyzing these effects important drivers that affect the customers’ decision making process are identified in an early stage. The gained information guides firms in their actions to reduce customers’ intentions to churn and prevent the actual churn. To analyze the effects the following research question is set:

To what extent do perceived costs and benefits influence churn intention and churn behavior for different groups of people in the Dutch Health insurance market?

From this main research question the following sub questions emerge: 1. What are the perceived switching benefits and perceived costs? 2. How do the perceived costs and benefits influence churn intention? 3. To what extent does churn intention lead to actual churn?

4. To what extent does churn intention acts as a mediator between the perceived costs and benefits and actual churn?

5. What are important segmentation dimensions? 6. How do the relevant segments look like?

7. To what extent do the different segments influence (moderate) the relationship between the perceived cost and benefits and churn intention?

8. To what extent do the different segments influence (moderate) the relationship between churn intention and actual churn?

(10)

10 past scientific literature. The hypothesis and the visualization of the conceptual model are also presented in the theoretical framework. In Chapter 3 the research design and methodology are outlined. The results of the analyses are addressed in Chapter 4 and to finalize this study Chapter 5 includes the discussion and conclusion.

2. THEORETICAL FRAMEWORK

In this section existing literature is studied with the aim to gain more insights in churn behavior. More information about underlying drivers is necessary to explain churn behavior. Therefore this section presents more information about churn, churn intention, the perceived benefits, costs and customer characteristics. Concluding this section the conceptual model with its hypothesis is visualized.

2.1 Churn

Nowadays, where competition is increasing and customers are more demanding, firms need to be more customer centric (Ascarza & Hardie, 2013). Churn management, part of customer relationship management (CRM), plays a large role in this and is highly important for firms. Overall churn refers to the fact that a customer ‘leaves’ a firm and goes to another firm. In service industries, like the Dutch health insurance market, there is a contractual setting involved. Churn in a contractual setting is defined as ‘the termination of a contract between the firm and its customer’ (Leeflang et al., 2015). In this situation churn is a binary issue, with the aim to explain churn or predict the likelihood that a customer churns in a given period. In the area of marketing several researchers developed models that explain or predict churn. Findings based on prior research also indicate the need for regular adaptation of churn models (Leeflang et al., 2015).

2.1.1. Importance for firms

(11)

11 to leave a firm, the firm loses revenue. If churn rates are very high the firms need to acquire new customers increases and directly leads to an increase in acquisition costs. More important in service industries, where a contract is involved, the loss of a customer also indicates the termination of (long term) relationship (Risselada, Verhoef & Bijmolt 2010). The loss of this valuable relationship comes with a loss in future cash flow trough cross selling or upselling. The importance of gaining more information about churn probabilities increases. When firms are in the possession of these probabilities the opportunity to anticipate to this increases. If a customer expresses a high churn probability the loss of a valuable relationship can be prevented by the firms’ actions. A way to gain insights in the churn probability is through churn intention.

2.1.2 Churn intention

For marketers it is important to gain more knowledge about the driving forces behind customer churn, in order to improve the ability to prevent or at least predict future churn. A common used proxy for behavior are customers’ self-reported intentions, in this study churn intention. In past literature churn intention is seen as easy-to-collect information to predict behavior. Churn intention is defined as ‘the customers’ self-reported likelihood to end a current relationship’, where actual churn is the definite termination of the contract or relationship (Wirtz et al., 2014). The distinction between the intention to churn and the actual churn is that intention refers to a hypothetical situation. Wirtz et al., (2014) indicate that in a hypothetical event customers tend to focus more on the outcomes, such as receiving a good price or better service. In contrary when the event approaches, customers focus more on the procedure (e.g. time and effort) to carry out the action. Previous research in the Dutch health insurance market confirms that the expression of a high churn intention more often leads to churn (Hendriks et al., 2009). The researchers also state that churn intention is an accurate predictor of churn in the Dutch market. From this can be assumed that churn intention exerts an important role in predicting churn behavior. Knowing more about the customers’ intentions is an opportunity to influence customer churn behavior. Based on these findings the following hypothesis is set:

Hypothesis1a: Churn is expected to be heavily influenced by the intention to churn. The stronger the customers’ intention to churn is, the more likely it is the customer actually churns.

(12)

12 churn intention is the customers’ attitude, not service quality and satisfaction. The best way to affect this attitude is to understand and manage the advantages and disadvantages, cost and benefits, of churn (Bansal & Taylor, 1999). Churn intention is seen as a direct predictor for actual churn, but churn intention may also act as a mediator. The SPSM model of Bansal & Taylor (1999) implicates this, in their study the effect of perceived switching costs is included. The results show that in some service industries it is more likely to see an indirect of the perceived costs on churn through churn intention. To indicate if this holds in the Dutch Health insurance market, the mediation effect is added. Churn intention mediates the relationship between the perceived benefits and costs and actual churn. In addressing churn intention as mediator, firms have the opportunity to indicate the probability of churn in an early stage by the level of churn intention a customer expresses. For this reason the following hypothesis is set:

Hypothesis1b: Churn intention acts as a mediator in the relationship between the perceived benefits/costs and actual churn

2.1.3 Churn in the Health Insurance industry

(13)

13 countries a similar tendency is noticed, prior research shows that older and less healthy customers experience higher switching barriers. Besides that, a main reason for switching to a different insurer in all countries is service quality. On the other hand price is not in all countries a reason for churn. In Belgium differences in price are negligible and therefore it is often not a reason for churn (Thomson et al., 2013).

The intention to churn is captured by factors that drive behavior and attitudes. Drivers for churn are extensively discussed in marketing literature. The next two sections indicate the perceived benefits and costs that influence churn attitudes.

2.2 Perceived benefits

Churning to a different health insurer could lead to benefits, these benefits are seen as drivers that lead to certain (churn) behavior. Churn benefits can be defined as ‘the perceived benefits customers experience from switching to a different insurer’ (Duijmelinck, Mosca & van de Ven, 2014). In order to find relevant perceived benefits, prior research is reviewed.

2.2.1 Price

Overall customers are sensitive to price, as mentioned in the introduction this was also one of the main reasons a large group of Dutch customers churned. Back in 1995 when Keaveny studied drivers for churn, price was already seen as the most prominent and sometimes the only reason for actual churn. In the health insurance industry price is be affected through the factors; premium, collective discount or special (transfer) offers/discounts. The insurance companies excess freedom to set their own prices. The differences in the basic insurance in the Netherlands are quite large, the prices range from the €92,00 up to €122,00 a month.1 According to Strombom, Buchmueller & Feldstein (2002) the sensitivity to health insurance premiums of customers is heterogeneous. One of the new trends known in the Dutch health insurance industry is the ‘budget policy’. The low premium of this policy might drive customers to a different insurer (Duijmelinck, Mosca & van de Ven, 2014). The Nederlandse Zorgautoriteit (Nza) conducted a research in 2016 to analyze the Dutch health insurance market, which contained a section about churn behavior. The research shows that the most important reason to churn to another insurer is based on the amount of the total premium. In 2016 customers saved up to 30,5 million euros on insurance premiums for only the basic insurance. The price sensitivity in this market was -0.7 in 2016, which indicates that a 1% increase in the premium

(14)

14 while the prices of the other insurance policies remain the same, there is a decrease in demand of 0.7% for that specific insurance policy. Also the research of Duijmelinck, Mosca & van de Ven (2015) highlight the important role of price related to the decision to switch. Not only in the Netherlands, but also in Germany and Switzerland this is the main reason for churn.

Another interesting phenomenon are the collectivity discounts, also known as ‘group contracts’. Insurance companies are allowed to provide discount up to 10% for customers who belong to a group. Most of the times this goes via the employer of the customers. The employer establishes an agreement with an insurance company to offer their employees (and family) a discount on their health insurance (Bolhaar, Lindeboom & van der Klauw, 2015). The aforementioned researchers conclude in their study that the presence of a group contract increases the probability to churn.

Next to the premium and the collective discount, insurance companies also compete on special offers to attract new customers. The insurance companies offer a discount when the premium is paid upfront for the entire year, ‘welcome gifts’ for new customers or a discount for customers who bring new customers to the insurer. These ‘welcome gifts’ can be in monetary form, but also materialistic for instance a tablet, an Xbox or a free travel insurance (Ministerie van Volksgezondheid, Welzijn en Sport, 2014). Based on the findings in prior research the following hypotheses are set:

H2a: The premium of other insurance companies is the most important perceived price benefit that drives churn intention.

H2b: The premium of other insurance companies is overall the most important perceived benefit that drives churn intention.

2.2.2. Service quality

(15)

15 handling claims, coverage decisions or the service of the customer center. The level of these aspects of service quality distinguishes a firm from its rivals. If a customer is not satisfied with the service quality at their current provider this could be a driving force to look into other options. The findings in prior research results in the following hypothesis for this research:

H3: The perceived benefit service quality positively affects churn intention. When the service quality of the current insurer is low, customer express a higher churn intention.

2.2.3. Package

The benefits that insurance packages offer affect churn intention and actual churn. First of all this can be through the coverage and compensation insurers offer. The insurances that are offered must include a package with the basic benefits set by the government. However, the new legislation since 2006 also came with some flexibility for insurers. Insurers have the opportunity to selectively contract with the several healthcare providers and design their own ‘products’ to better attract customers (van den Berg et al., 2008). In most cases these ‘designed products’ are insurance policies with limited conditions at a lower price, but offer a lower compensation in non-contracted health care. Besides that, there is a limited network of contracted healthcare providers (Nza, 2016). Secondly, in addition to the basic insurance, customers can top this with a supplementary insurance. The influence of this extra benefit is often underestimated. Taking into account the Dutch health insurance market the importance of this benefit is noticed, almost 84% of the customers have a supplementary insurance. In 2013 this was one of the main reasons customers churned (Duijmelinck, Mosca & van de Ven, 2014). The benefits of the supplementary insurance differ per insurer and the jointly purchase of a basic and a supplementary insurance might lead to extra benefits. Previous research of Wieringa & Verhoef (2007) also indicate that additional services, for instance supplementary insurance, might influence churn. In line with this previous research the following hypothesis is set:

H4a: The perceived benefit (better) coverage at other insurance companies positively influences churn intention.

H4b: The positive effect of supplementary insurance on churn intention is stronger than the effect of coverage.

(16)

16

2.3 Perceived costs

Prior research suggests that perceived costs are key determinants for churn behavior (Bansal & Taylor, 1999). When a firm wants to manage switching costs it is important to know which costs customers perceive. Switching costs are defined as ‘the costs that customer experience with the process of switching from one firm to another’ (Burnham, Frels & Mahajan, 2003). In order to gain more knowledge about churn behavior in the Dutch health insurance industry it is important to represent relevant perceived switching costs. These costs may have monetary consequences, but can also cause inconvenience and effort. Prior research shows that monetary switching costs are more related to churn intention and non-monetary to actual churn (Wirtz et al., 2014). Burnham, Frels & Mahajan (2003) suggest three higher-order types of switching costs: procedural, financial and relational.

H5: The higher the overall perceived switch costs, the lower customers’ intention to switch.

2.3.1. Procedural loss

Procedural switching costs primarily focus on the expenditure of time and effort, which are non-monetary costs (Burnham, Frels & Mahajan, 2003). In the health insurance industry there are a lot of comparable products which information needs to be analyzed to make a decision. Customers may experience this as overwhelming and experience difficulties in making the right decision. Besides that, when customers made their decision they will be confronted with certain learning costs and setup costs. Time and effort need to be invested to learn the procedures and rules from the new insurer (Duijmelinck, Mosca & van de Ven, 2014). The switch to a different insurer also comes with some economic risk costs. When a customer has insufficient information about the new insurer the uncertainty that this may lead to a negative outcome needs to be accepted (Burnham, Frels & Mahajan, 2003).

H6: Procedural losses negatively affects churn intention. The higher the losses customers’ experience, to lower the intention to churn.

2.3.2 Financial loss

(17)

17 not apply to new customers (Burnham, Frels & Mahajan, 2003). These benefits also indicate the loss of favorable conditions that are offered in the customers’ supplementary insurance. Especially for customers with health issues this plays an important role (Duijmelinck, Mosca & van de Ven, 2014). The uncertainty of monetary loss costs are more related to other churn intention that other non-monetary costs. Customers first focus on outcome-related variables and therefore not even consider or underestimate the effect of the non-monetary costs (Wirtz et al., 2014). The loss of monetary costs indicate in most cases onetime financial expenses that occur from churning to a different insurer (Burnham, Frels & Mahajan, 2003). The new rules since the reform, where selective contracting is allowed, lead that churning to a different insurer also leads to changing the medical providers. Based on these prior research certain assumptions can be stated, therefor the following hypothesis are set:

H7: Financial losses are the most important costs that negatively drive churn intention.

2.3.3. Relational loss

Finally customers can experience relational switching costs, which indicate personal and brand relationship loss costs. After a long relationship customers become familiar with an insurer and its rules. The loss of this relationship leads to a loss of identification and comfort which is not directly created with a new insurer. Relational costs can therefore lead to emotional or psychological discomfort (Burnham, Frels & Mahajan, 2003). This is also in line with the research of Duijmelinck, Mosca & van de Ven (2014) who argue that these costs may lead to irrational behavior, for instance status quo bias. Besides that these scholars also find evidence that the longer the relationship lasts, the less likely customers are in churning. In line with these findings the following hypothesis is set:

H8: Relational losses are the least important perceived costs that negatively affect churn intention.

The perceived benefits and costs may differ per customer or customer group. For this reason the effect of customer characteristics are taken into account in the next section.

2.4 Segmentation based on customer characteristics

(18)

18 an effect on the costs and benefits, on churn intention and on actual churn behavior. Especially in health care customers are very heterogeneous. Research of Verhoef and Wieringa (2007) shows that there are different customer segments in the Dutch energy market, which all have different churn intentions. It does not necessarily mean that customers with high churn intentions actually churn, but it gives a direction that there are heterogeneous groups. In this study the effect of segmentation based on customer characteristics is added. In the next paragraphs the customer characteristics that are analyzed for the segmentation in this study are presented. During the selecting process of these variables current knowledge is taken into account, but also the need create a full customer profile.

2.4.1 Demographic characteristics

Demographic characteristics are one of the most important characteristics used in segmentation. The main reason for this is that customer behavior is most likely to vary closely with these characteristics (Kotler, 1980). A first characteristic of interest is age, the older the people are the more satisfied they are (Bryant and Cha, 1996). This could be due to different norms and values or level of experience. Also level of income plays a role in churn behavior. In the Dutch health insurance market, people with a low income receive a monthly subsidy. The subsidy can reach a maximum of €88 each month, which is quite a large amount. Customers with a low income are overall more price sensitive than customers with a higher income (Duijmelinck, Mosca & van de Ven, 2014). When these customers found an insurance that fits their needs, they are less likely to churn. This could be due to several reasons, for instance unexpected costs. A characteristic that is in most cases related to income is education, this is also indicated in results of prior research. Customers with a high education level express a higher churn rate (Duijmelinck, Mosca & van de Ven, 2014). This could be due to the fact that high educated customers always look for the best option, spend more time and effort in their search which lead to better option. Another reason could be the level of interest or a financial advisor who takes care of the insurances. Also household size may affect churn behavior. In the study of Risselada, Verhoef & Bijmolt (2014) the researchers state that household size is a variable which is relevant in predicting churn behavior. To indicate if the effect of household size is also relevant in the Dutch health insurance market, this variable is added in the segmentation analysis. Based on this prior research the following hypothesis are set:

(19)

19 H9c: Segments including young high educated customers are more likely to churn.

H9d: Segments including young high educated customers have a higher churn intention.

2.4.2 Geographic characteristics

The aim of segmentation based on geographic characteristics is to divide groups of customers into geographical units (Kotler, 1980). In this study the focus is on the Dutch market, therefore the region a customer lives in is added. Taking into account the degree of concentration in the Dutch market in 2016, quite some differences occur (Nza, 2016). A high degree of concentration is seen in rural areas for example Friesland (highest) and Zeeland. The high degree of concentration indicates that there are only a few insurers and these few insurers have a high market share. The lowest degree of concentration is in Zuid Holland followed by Utrecht. These areas are characterized as urban areas. When a low degree of concentration occurs there are many small insurers involved, which leads to more competition and more choices for the customer (Nza, 2016). Therefore, the fact that a customer lives in a rural are or a more urban area may affect churn intention or actual churn. Besides the area, the size of the community can play a role. In a small community customers know each other and have more social contact. The likelihood neighbors know about each other insurers is larger in a small community. By knowing this, firms can focus more on the geographical needs and behavior of customers (Kotler, 1980).

H10a: The level of churn intention is higher in urban areas than rural areas.

H10b: Segments including customers living in an urban area are more likely to churn. H10c: The region a customer lives in determines the type of insurer (national vs regional). H10d: Customers living in the western part of the Netherlands are more likely to churn.

2.4.3. Channel usage

(20)

20 but also state that there is a large difference for product and services (Neslin et al., 2006). For this reason these are also taken into account in this study.

H11a: Customers that search online are better informed and experience a high churn intention. H11b: Customers that use online channels to gather information are less likely to churn.

Now that more information about the individual segmentation variables is known, the moderating effect of the segments is taken into account. Berget et al. (2002) indicate that the homogenous subgroups that are formed in the segmentation analysis experience the benefits and costs in a similar way. Eventually this lead to similar behavior, in this study churn intention or even actual churn. For this reason it is expected that the homogenous subgroups differ from each other. To analyze if that is the case in the Dutch health insurance industry the following hypothesis are set:

H12: The effect of the perceived benefits and costs on churn intention is different per segment. H13: The effect of churn intention on churn is different per segment

(21)

21

2.6 Conceptual model

The conceptual model represents, based on the literature review, a visualization of the relationship between the variables and its hypothesis (figure 1). The model assumes the effect of the perceived benefits and costs on churn is mediated through churn intention. Next to the mediating effect the moderating effect of segmentation is added on both of the relationships. Therefore it is concluded that the model deals with moderated mediation.

FIGURE 1: Conceptual model

The research question that is set in the introduction is: To what extent do perceived costs and benefits influence churn intention and churn behavior for different groups of people in the Dutch Health insurance market?

(22)

22

3. METHODOLODY

In this section the methodology part of this study is discussed. The first paragraph discusses the characteristics of the dataset, followed by a short description of the selected variables and the data preparation steps.

3.1 Dataset

In this study a database with customer data is used to analyze churn behavior in the Dutch health insurance market. In total there were five separate datasets, one dataset with purchases, one with background characteristics, one with offline orientation information, one with search information and finally one with website information. From this five subsets, one set is created. Figure 2 provides an overview with the information from the five sets transformed to the final dataset used in this study.

FIGURE 2: Overview of subsets with relevant information

(removed due to confidentality)

(23)

23 households. Each household is marked with a unique ID number. In the dataset three phases are distinguished, the so called S-phase where customers highlight their intentions and the aspects that influence their decision. The O-phase where customers indicate how the necessary information is obtained and finally the P-phase where the final decision is made. The final decision to churn or stay is subdivided in churn, active extender and passive extender. Customers that decide to churn to a different insurer are defined as actual churners.

3.1.1 Restructuring the dataset

The complexity of the dataset is increased by the number of repeated measure moments. As mentioned before, there are 21.324 cases in total. One household contains multiple cases, where each row is one time point. The maximum time points a household has is four, each month one measure moment, which is also defined as a ‘long’ data format. In order to analyze the data correctly and gain a better overview, a ‘wide’ format is created. In the ‘wide’ format the selected cases are restructured in one case (row) for each household. Which leads to an aggregated dataset based on household level. The variables ID number, age, income, household size, district size, district, stratum and education level are defined as identifier variables. The information of each identical household is now captured in one row and each response is in a separate column. The main reason the data is transformed to this so called ‘wide’ format is due to the fact that data in this format is easier to analyze. Different types of analysis require a different set up of the data, in this study the ‘wide’ format is most appropriate. Restructuring the dataset into this new format comes with a lot of extra variables. The next section indicates the transformation of these variables.

3.2 Variables

Initially the dataset consisted of five different subsets with all their own subject and information. In order to answer the research questions in this study, a new dataset with all necessary information and variables is created. Non-relevant variables are deleted and not taken into account. A full list of the variables used in this study is provided in Appendix A, table A1.

3.2.1 Transforming and recoding variables

(24)
(25)

25

TABLE 1: Overview of recoded variables Variable Old values New values

Age String (18 to 113) Categorized in groups 1: 18 – 25 4: 46 – 55 2: 26 – 35 5: 56 – 65 3: 36 – 45 6: 66 – 115 Income Category with 22 groups

Categorized down to 5 groups

1: No income 4: €2500 tot €3500 2: tot €1500 5: €3500 of meer 3: €1500 tot €2500 Education Category with 15 groups

Categorized down to 7 groups

1: No education/does not apply 5: Middle level applied education (MBO) 2: Special education 6: Higher professional education (HBO) 3: Elementary education 7: University (WO)

4: Secondary education Household size * String (1 to 13) Categorized in groups 1: Small (1 to 2 persons) 2: Medium (3 to 4 persons) 3: Large (5 or more persons)

District Category with 5 groups

Categorized in 4 groups

Three big cities and rest west are put together * Based on CBS data of 2013

The table shows that household size and age in the original dataset are defined as string variables. All the other socio-demographic variables are categorical, therefore it seems logic that these variables are categorized. Besides that, the interpretability is improved and there is a better overview of the size of each group. The variable income is minimized from 22 categories to 5 categories. After checking several distributions, this division seems most appropriate. Finally, several variables in the original dataset are string variables that contain text, not one word but complete sentences. In some cases these string variables are an additional explanation of previous answers in the numeric variables and in other cases the variable ‘Other’ is clarified. Especially in the last case correctly recoding this information is highly important. In order to improve interpretation and increase simplicity all string variables are manually recoded. The main reason manually recoding is applied is due the ambiguity of results. Either the textual information is deleted due to the fact that there is already an answer indicated in the numeric variables or the information is recoded in the existing ‘Other’ variable. New labels are created for frequent responses to prevent information loss. A full list of the recoded string variables is provided in Appendix A, table A2.

3.3 Missing variables and oddities

(26)

26 mentioned two final decisions, which makes this observation unreliable and therefore this household is also deleted from the dataset. To check if deleting these households effects the representativeness of the dataset, an ANOVA test is performed. The results show that there is a statistically significant difference, mainly caused by age, between the groups for some of the variables. (Table A3 in Appendix A). For this reason the distribution of these significant variables is analyzed. In all cases the results of the Kolmogorov-Smirnov normality tests where significant. From this can be concluded that the deleted as the non-deleted households significantly deviate from a normal distribution. To rectify this, weighting is applied for the remaining 5355 households.

3.3.1 Weighting

In order to obtain a representative and generalizable dataset for the Dutch health insurance, weighting for the variable age is applied. The younger age group (18 up to 45) is underrepresented. There could be several reasons for this, one of these reasons is that parents administer the insurance details of their children or that younger people are less willing to participate in panels. The largest age group of the sample before weighting is the group in the age 56 – 65 years old, where the smallest group is 18 – 25 years old. Based on data from the Dutch ‘Centraal Bureau voor de Statistiek’ (CBS), data of the population in 2013 is extracted and used for weighting the sample. Taken into account the observed data in our dataset and the actual age distribution for the Dutch population in 2013 a weighting factor is derived. The results after weighting, described in table 2, are more equally distributed and therefor the dataset is better represented by the different age groups.

TABLE 2: Distribution of age before and after weighting

Age group Before weighting After weighting

18 - 25 2,3% 12,6% 26 - 35 11,4% 15,2% 36 - 45 18,1% 17,7% 46 - 55 22,9% 18,8% 56 - 65 24,5% 16,2% 66 - 115 20,8% 19,5% 3.4 Churn analysis

(27)

27 3.4.1 Factor analysis

A first step is factor analysis, this type of analysis checks the relationship between a set of interdependent variables. The main goal of this procedure is to reduce or summarize the data. Underlying ‘factors’ are examined that help explain the correlations between a set of variables (Malhotra, 2008). In this study factor analysis is applied to the perceived benefits and the perceived costs. To assure that factor analysis is appropriate for these variables, a correlation matrix is constructed. The results from the correlation matrix, the Bartlett’s test of sphericity and the Kaiser-Meyer-Olkin (KMO) measure of sampling adequacy identify the appropriateness. A common method of factor analysis is principal components analysis (PCA), this type of analysis considers the total variance in the data (Malhotra, 2008). To define the underlying construct in the data PCA with Varimax rotation is applied. The Varimax procedure defines variables with high loadings on a factor and minimizes these, this results in factors that are not correlated (Malhotra, 2008). According to Wieringa and Verhoef (2007), the advantage of PCA is that multicollinearity is reduced in subsequent analysis due to the use of orthogonal factors.

3.4.2 Logistic regression analysis

A second step is to indicate the separate effects and significance of the variables. Linear and logistic regression analysis support this. Regression analysis is a procedure that analyzes the associative relationship between variables. This type of analyses explains whether a relationship exists, the strength of the relationship, determines the structure or form of the relationship and predicts the values of the dependent variable (Malhotra, 2008). Logistic regression is applied to determine the effect between churn intention and actual churn. Churn is a binary variable where 0 = stay and 1 = churn. Due to this binary effect linear regression is not appropriate. Logistic regression, also called the binary logit model, measures the probability of an event taken place. By performing this logistic regression customer churn is explained based on churn intention. Also the relationship between the independent variables and churn intention are analyzed by performing a logistic regression analysis due to the fact that churn intention is recoded in a dummy variable.

3.4.2 Moderation analysis

(28)

28 and direction of the effect of the moderator. To determine this the following models are estimated;

 Y is a function of X  Y is a function of X and z  Y is a function of X, Z and X*Z

If there is a significant parameter for X*Z a moderator effect is indicated. In order to measure the effect of the moderator on both of the relationships, this will be performed twice. Firstly, to test the moderating effect of segmentation on the relationship between the perceived benefits/costs and churn intention. Secondly, to test the moderating effect of segmentation on the relationship between churn intention and churn

3.4.3 Mediation analysis

Moderation and mediation are often confused with each other. Mediation accounts for the relation between independent (X) and dependent variable (Y). There is an indirect effect, the mediator reflects the process or mechanism between X and Y. The difference between a moderator and a mediator is that the moderator specifies when an effect will hold. In contrary, the mediator specifies the ‘how’ and ‘why’ a certain effect occurs (Baron & Kenny, 1986). To clarify this, a graphic representation is shown in figure 3.

FIGURE 3: Graphical representation mediation

(29)

29 mediator, in example path a. This also accounts for the levels of variations between the mediator and the dependent variable, path b. In the case that both paths are controlled for it is assumed that the relationship between the independent and dependent variable is not significant anymore, with the strongest mediation occurs when the path c is zero. (Baron & Kenny, 1986).

The researchers define four steps to establish mediation:

1. The independent variable must affect the dependent variable (path c) 2. The independent variable must affect the mediator (path a)

3. The mediator must affect the dependent variable (path b)

4. The effect of the independent variable on the dependent variable controlled by the mediator must be zero (path c’)

If all the steps are met than the conclusion that the mediator, in this case churn intention, completely mediates the relationship between the independent and dependent variable. If only the first three steps are met, partial mediation is the conclusion (Baron & Kenny, 1986). According to Kenny (2016) are the steps for testing mediation the same, regardless which method of estimation is used. Several logistic regression analysis are performed to indicate the effect of the different paths. The syntax Nathanial Herr provided is used as a guidance (see Appendix C).

3.7 Segmentation analysis

(30)

30 and is, according to Wieringa & Verhoef (2007), a powerful method. By performing this analysis valuable insights into churn, the drivers and the differences between the segments is obtained.

4. RESULTS

In this section the results of the several analysis are presented. The first part describes a general view of the dataset, some descriptive statistics are shown to get to know the data. Hereafter, the results of the basic regression analysis is shown followed by the results of the more complex analysis.

4.1 Descriptive statistics

4.1.1 Demographic and geographic statistics

After restructuring the dataset, the total sample consists of 5355 individual households. The age of the respondents vary between 18 and 113 years, with an average age of 48,14 years (SD=16,54). The largest group of respondents is in the age groups 66 years and older and the smallest group is in the youngest age group 18 – 25 years old, see table 3.

TABLE 3: Overview demographic characteristics

The largest group of respondent has an income between €1500 and €2500 each month. The other groups are quite equally divided, a difference is that the highest income level is covered by only 10% of the respondents. Most of the respondents (31,8%) finished a middle level applied education and the most common household size is small which includes 1 or 2 persons. Next to the demographics, there is also geographical information about the respondents. The

Variable Category Percentage

Age 18 - 25 12,6% 26 - 35 15,2% 36 - 45 17,7% 46 - 55 18,8% 56 – 65 16,2% 66 plus 19,5%

Income No income/does not want to disclose 7,7%

Up to €1500 24,3%

€1500 till €2500 34,4%

€2500 till €3500 23,1%

€3500 or more 10,4%

Education level No education/does not apply 0%

Special education 0,6%

Elementary education 1,8%

Secondary education 29,5%

Middle level applied education (MBO) 31,8% Higher professional education (HBO) 25,3%

University (WO) 11%

Household size Small 70,5%

Medium 23,8%

(31)

31 district in which the respondents live in and the size of the district is known. An overview of these variables is shown in table 4.

TABLE 4: Overview geographic characteristics

Most of the households (38,9%) live in the Western part of the Netherlands and the smallest group (11,7%) represents the North of the Netherlands. Taking into account the actual numbers per region in the Netherlands in 2013, similar effects are noticed.2 Most people live in the Western part of the Netherlands. The size of the district is also taken into account, almost 36% of the respondents live in a district with 20.000 up to 50.000 inhabitants, followed by 30,6% that lives in a district with more than 100.000 inhabitants. These results can be linked to the fact that most respondents live in the West of the Netherlands where the larger cities are located. The actual numbers in 2013 in the Netherlands indicate that the dataset is an appropriate representation of the reality in 2013. 3

4.1.2 Channel usage

In the analysis of the channel usage the following remarkable fact arose. Of the 5355 respondents in the dataset, only 2293 respondents answered questions concerning channel usage. The rest of the respondents, more than half, did not give any information about their channel preferences. If this variable is taken into account in the segmentation analysis a lot of information, of 3062 respondents, will be lost. Due to this large amount of information loss the decision to leave out the variables concerning channel usage is made. The focus in the segmentation is on the demographic and geographic characteristics.

4.1.3 Churn intention and actual churn

In this study the relationship between churn intention and actual churn plays an important role. To gain more information about these variables, some descriptive statistics are shown. The variable churn intention is classified in four categories. The largest group of respondents shows

2 http://cbs.overheidsdata.nl/70634ned

3 http://statline.cbs.nl/Statweb/publication/?DM=SLNL&PA=82000ned&D1=48,51&D2=a&HDR=T&STB=G1&VW=T

Variable Category Percentage Actual numbers in 2013 in the Netherlands

District North 11,7% 10,2%

East 23,4% 21,2%

South 26% 21,4%

West 38,9% 47,2%

Size of district Up to 20.000 inhabitants 12% 11,1%

(32)

32 a no intention (47,9%), 24,2% indicates a medium intention to churn, 3,4% shows a high intention and the remaining part of 24,6% indicates to have a low intention. Table 5 shows an overview of the distribution between the variables churn intention and churn

TABLE 5: Crosstab of churn intention and actual churn

Switcher Active extender Passive extender Total

No intention % within intention 5,0% 3,7% 91,4% 100,0%

% within type of buyer 14,2% 36,8% 55,8% 47,9%

% of total 2,4% 1,8% 43,8% 47,9%

Low intention % within intention 8,4% 3,2% 88,4% 100,0%

% within type of buyer 12,3% 16,6% 27,6% 24,6%

% of total 2,1% 0,8% 21,7% 24,6%

Medium intention

% within intention 39,6% 8,7% 51,6% 100,0%

% within type of buyer 57,2% 44,3% 15,9% 24,1%

% of total 9,6% 2,1% 12,5% 24,1%

High intention % within intention 80,1% 3,3% 16,6% 100,0%

% within type of buyer 16,3% 2,4% 0,7% 3,4%

% of total 2,7% 0,1% 0,6% 3,4%

Total % within intention 16,7% 4,8% 78,5% 100,0%

% within type of buyer 100,0% 100,0% 100,0% 100,0%

% of total 16,7% 4,8% 77,8% 100,0%

The churning respondents overall expressed a high or medium intention, which seems logical. In contrary the passive extenders mostly no intention. Remarkable is the fact that there are some respondents who churned, expressed to have a low or even no intention. An assumption that can be made is the fact that these respondents already churned and did not define their intention. A large group of churned respondents expressed a medium intention, which is higher than the group with a high intention. This could indicate that respondents experience some doubt in the churn process and are later on persuaded.

4.1.4 Perceived costs

In the analysis of the perceived costs the fact arose that the variables that define the perceived cost are only submitted to a select group of respondents. Unfortunately only the responses of the passive extenders to these variables are in the dataset, therefore analyzing these variables does not obtain any information about the respondents that churned. For this reason, the perceived costs are left out of the analysis. If they would be taken into account biases might occur and the results are no longer representative for all the respondents.

4.1.5 Perceived benefits

(33)

33 to very important role. Also the effect of a supplementary insurance is taken into account, due to the fact that this is marked as a main reason for churn. Unfortunately, due to limitations in the dataset it is not possible to analyze the effect of this variable. Only a small group of 1184 respondents indicated the type of insurance purchased. This comes with another downfall, the question is asked in the P-phase. At this moment respondents already made the decision to churn or stay. Therefore this variable only indicates the decision that is made and does not provide any information about the influence on their decision. To gives some insights in the presence of a supplementary insurance are indicated in table A4 in the appendix A. The results confirm indeed the fact that most respondents have a supplementary insurance.

An overview of the most remarkable results is shown in table 6. The variables with the most important role are; lower premium at a different insurer and a better coverage and compensation at a different insurer. Another remarkable fact is that some of the variables play an important role for a large group but also for a large group no role. Next to show the last three variable not that important role. The least important variable is the fact that respondents experience difficulties in paying the premium for their health insurance.

TABLE 6: Overview of most important results perceived benefits

Switcher Active extender

Passive extender Premium of current insurance is to high Important role 35,1% 27,5% 76,8%

Moderate role 25,8% 28,3% 77,4%

Premium increases (too) much in 2014 Important role 32,5% 32,8% 30,7%

No role 30,1% 27,9% 30,1%

Lower premium at different insurer Very important role 37,8% 27,9% 21,9%

Important role 39,4% 29,5% 34,8%

Dissatisfied with coverage/compensation Important role 27,2% 26,2% 28,2%

No role 34,1% 38,1% 38,9%

Better coverage at different insurer Very important role 29,6% 22% 21,7%

Important role 45,2% 38% 38,2%

Attractive collective discount Important role 34,6% 34,8% 33,7%

No role 26,5% 31,3% 33%

Interesting offer a different insurer Important role 35,4% 33,6% 29,6%

No role 23,5% 25% 31,1%

Difficulties in paying current insurance Moderate role 20,2% 15,9% 19,3%

No role 52,8% 62% 52,1%

Dissatisfied with less contracted health care

Moderate role 21,9% 26,6% 21,4%

No role 47,9% 40,2% 44%

Dissatisfied with the service quality of the current insurer

Important role 19,5% 19,7% 21,9%

No role 51,8% 58,6% 53,8%

4.2 Factor analysis

(34)

34 multicollinearity, that may influence further analyses are resolved. Factor analysis is usually applied for the independent variables, in this study this are the perceived benefits and costs. Unfortunately the perceived costs can’t be analyzed due to limitations in the dataset which will be discussed later. For this reason the factor analysis only indicates the perceived benefits. A first check of the factor analysis is the correlation matrix, see table 7. A value large than 0,5 indicates that the correlation between the pairs of variables can be explained by other variables. All values are under 0,7 are noted as medium correlations, from 0,7 – 0,9 are high correlation and 0,9 to the maximum of 1 are high correlations.

TABLE 7: Correlation matrix perceived benefits

In the matrix the values larger than 0,5 are highlighted, in total 11 variables. There are relative high correlations between the premium variables (PREM), but also between premium and coverage (COV) and coverage and service quality (SQ). A second step is to indicate if factor analysis is appropriate, to measure this the principal component analysis is executed. The Kaiser-Meyer-Olkin (KMO) measure is 0,903 which is larger than 0.5 and therefore it is likely that the data ‘factors’ well. Also the Bartlett’s test of sphericity is significant, the null hypothesis that the correlation matrix is an identity matrix is rejected. The results of the principal component analysis are shown in table 8. There are several criteria to obtain the right number of factor. First, the eigenvalue, the variance a factor explains needs to be larger than 1 (blue). Secondly, the total explained variance (green) needs to exceed 60% and thirdly the factors need to explain 5% each (orange). The scree plot (figure 4) also gives a good indication of the appropriate number of factors. Ideally this is a steep curve followed with a bend and then a horizontal line. In this scree plot this pattern is seen, the best fit is the number of components before the point that starts the flat line. From this can be concluded that two factor seems the best solution. Due to the difficult interpretability of the unrotated factor matrix, orthogonally rotation with VARIMAX is applied. In table 9 the rotation results are shown in two components,

PREM1 PREM2 PREM3 COV1 COV2 COLL SPEC PREM4 COV3 SQ

(35)

35 the value of the loading should exceed 0.5. The first and the second variables are close to each other, but the remaining variables show clear differences.

TABLE 8: Results factor analysis TABLE 9: Rotated component matrix

FIGURE 4: Scree plot

Looking at the components and the variables in the components the distribution is not as expected from the literature review. The results from the factor analysis divide the variables in negative and positive benefits. The negative variables indicate the dissatisfaction at the current insurer and the positive variables indicate the gains respondents receive by another insurer. A disadvantage of forming the factors is that the underlying constructs, price, coverage and compensation and service quality are not taken into account. Without the factors the overall Cronbach’s Alpha value is 0,883 which is high, this doesn’t improve by deleting one of the items. In order to maintain the underlying constructs, the factors are not used in further analysis in this study.

Initial Eigenvalues

Total % of Variance Cumulative %

(36)

36

4.3 Segmentation analysis

In order to test to hypothesis the segmentation analysis is needed and therefore discussed first. As mentioned in the methodology part a Latent Class Analysis (LCA) is performed to divide the sample into relevant clusters. Normally a valid number of models to start with is 1 to 4, due to the fact that this no led to the adequate results this is increased to 10 models. The output of the number of models is shown in table 10.

TABLE 10: Results Latent Class Analysis (LCA)

The BIC statistic is analyzed to indicate which model fits best. The model is preferred according to the lowest BIC value. The results show model five is the model with the best fit, so the sample is divided into 5 different segments. The size of the segments is indicated in table 11. From this can be seen that segment 1 is the largest segment and 5 the smallest.

TABLE 11: Segment size

Segment Number Percentage

Segment 1 1810 respondents 33,8% Segment 2 1509 respondents 28,2% Segment 3 1130 respondents 21,1% Segment 4 673 respondents 12,6% Segment 5 232 respondents 4,3% Total 5355 respondents 100%

In the following part each of the segments are described and given a name that profiles their characteristics.

Segment 1: Retirees

(37)

37 Netherlands, but also quite a large group (28,5%) lives in the South. The district size is characterized by 20.000 up to 50.000 inhabitants.

TABLE 12a: Overview segment 1

Segment 1 % in segment % of total dataset

Age group 66 plus 42,6% 14,4%

Income 1500 – 2500 euros 42,3% 14,3%

Education Secondary education 59,3% 20%

Household size Small 97,5% 33%

District West 40,2% 13,6%

District size 20.000 – 50.000 35,8% 12,1%

Segment 2: Middle class

Most of the respondents (35,7%) in this group are in the age category 36 – 45 and 32,1% in the age 46 between 55 years. The monthly income in this segment is between 2500 and 3500 euros. The household size in this segment is medium, 2 or 3 persons. The majority in this segment lives in a district with 20.000 to 50.000 inhabitants and mainly in the Western part of the Netherlands. The largest group of respondents (39,6%) in this segment have followed a middle level applied education. The relative high income could be due to the experience these respondents build over the years.

TABLE 12b: Overview segment 2

Segment 2 % in segment % of total dataset

Age group 36 – 45 35,7% 10%

Income 2500 – 3500 euros 42,3% 10,1%

Education Middle level applied 39,6% 11,2%

Household size Medium 80,5% 22%

District West 36,2% 10,2%

District size 20.000 – 50.000 42,6% 12%

Segment 3: Upcoming starters

(38)

38

TABLE 12c: Overview segment 3

Segment 3 % in segment % of total dataset

Age group 18 – 25 35,5% 7,5%

Income 1500 – 2500 euros 42,9% 10,1%

Education Middle level applied 43,7% 9,2%

Household size Small 99% 20,9%

District West 43,1% 9,1%

District size 100.000 or more 37,2%% 7,8%

Segment 4: Wealthy elderlies

The respondents in segment 4 are in the age group 66 years and older. The income level is 2500 - 3500 per month. All respondents live in a small household with 1 or 2 persons, which relates to the age group. In most cases the respondents in this age group are with a partner or alone. The district size is 20.000 up to 50.000 inhabitants in the Western part of the Netherlands. The education level is in most cases a higher professional education (HBO), which is related to the income level

TABLE 12d: Overview segment 4

Segment 4 % in segment % of total dataset

Age group 66 plus 40,6% 5,1%

Income 2500 - 3500 euros 38,6% 4,9%

Education Higher prof. educ. 68,6% 8,6%

Household size Small 98,7% 12,4%

District West 42,5% 5,4%

District size 20.000 – 50.000 36,4% 4,6%

Segment 5: High educated youngsters

Segment 5 is characterized by the young respondents in the age group 18 – 25 years old, most of the respondents earn an amount up to 1500 euros each month. Logically, due to the fact that this is a young age group, the household size is also small. A large group of the respondents (49,6%) lives Eastern part of the Netherlands in a district with more than 100.000 inhabitants. A remarkable fact is that 72% attended or still attends an education at the university.

TABLE 12e: Overview segment 5

Segment 5 % in segment % of total dataset

Age group 18 – 25 95,7% 4,2%

Income Up to 1500 euros 79% 3,4%

Education University 72% 3,1%

Household size Small 97,8% 4,2%

District East 49,6% 2,1%

(39)

39

4.4 Regression analysis

4.4.1 Hypothesis 1a: Effect of churn intention on actual churn

In order to analyze the effect of churn intention on actual churn a logistic regression (logit) analysis is performed. As mentioned a dummy variable for churn intention is created to create a more powerful model. The Omnibus test p-value, which indicates if the model fits the data is highly significant (p-value = .000 < .01). To indicate how well the model fits the data, the McFadden R2 value is calculated and the remaining pseudo R2 measurements are interpreted. The McFadden R2is 0.07, which is quite low. In the research of Wieringa and Verhoef (2007) a McFadden R2 of 0.25 is found reasonable. The Nagelkerke R2 is 0.114 and Cox Schnell R2 is

0.068, the latter is not that often used due to the fact that it will never reach the value one (Leeflang, 2015). Overall the explained variance in the model is between 7% and 11%. Furthermore the results of the classification table indicate that the percentage of correct classifications is 85,3%. The remaining results of the logistic regression are in table 13.

TABLE 13: Logistic regression analysis

Variable B S.E. Wald Sig. Exp(B) 95% C.I.for EXP(B) Lower Upper

High intention (dummy) 3,168 ,190 277,873 ,000 23,748 16,364 34,465

Constant -1,775 ,040 2003,438 ,000 ,169

Referenties

GERELATEERDE DOCUMENTEN

The results of our previous analysis (individual effects model) indicate that an increase in change in activity (i.e., higher changes within sessions), decreases customer churn..

Given the different characteristics of the online and offline channel, and the customers that use a respective channel, channel choice is expected to moderate the

Theoretical Framework Churn Drivers Relationship Breadth H1: - Relationship Depth H2: - Relationship Length H3: - Age H4: - Gender H5: - Prior Churn H6: + Price H7: + Promotion H15:

›  H4: Average product price positively influences the effect of the amount of opens on customer churn.. ›  H5: Average product price positively influences the effect of the amount

Adding a social influence variable and historical data to the model, resulted in highly significant, strong beta’s which influenced the predictive power of the churn model in a

H1b: Churn intention acts as a mediator on the relationship between the perceived benefits/costs and actual churn Accepted (Mediation) H2a: The premium of other insurance companies

To identify interaction effects that can have a moderating effect on the drivers of churn, a Pearson Chi-square correlation test has been performed for the variables of

Multiple variables have been added as moderators on the effect of perceived price on churn: customer dissatisfaction, a factor for the different insurers, the usage of