Factors for a Healthy Customer Base: Finding Key Factors of Customer Churn in the Dutch Health Insurance Industry

Factors for a Healthy Customer Base:

Finding Key Factors of Customer Churn in the Dutch Health Insurance Industry

By

Ruben van Brug

2

Factors for a Healthy Customer Base:

Finding Key Factors of Customer Churn in the Dutch Health Insurance Industry

Master Thesis

By

Ruben van Brug

University of Groningen Faculty of Economics and Business

MSc Marketing Intelligence January 21, 2016

1st supervisor: Dr. Pr. T.H.A. Bijmolt

2nd supervisor: Prof. Dr. Ir. K. van Ittersum

IJsselstraat 39 9725 GB Groningen

(+31)6 21 24 76 73 r.c.van.brug@student.rug.nl

3

Management summary

Loyalty programs have received massive research attention in previous literature (e.g. Bijmolt et al., 2010). Even today, loyalty programs are extensively studied around the world. Loyalty programs are known for their positive effect on customer loyalty and for the financial benefits to the firm (Mimouni-Chaabane and Volle, 2010). This research, in contrast to most previous studies, is not targeted at frequency reward programs (where you buy more to get more), but instead is targeted at a loyalty program which rewards customers for living a healthy lifestyle. Loyalty programs aimed at changing one's behavior result in customers who feel better, live healthier, and on the long-haul become less healthcare consuming. This positive behavioral change is rewarded with points, which can be redeemed for rewards.

This study looks into the factors which determine customer loyalty in the Dutch health insurance industry. For this study, data from a major Dutch health insurer, Menzis, is used. Menzis has a loyalty program, SamenGezond, which encourages customers to live more healthy. SamenGezond rewards customers when they live a healthy lifestyle (e.g. by exercising or eating healthy). To assess the loyalty program's effectiveness, key drivers were studied to identify the factors which mainly drive customer loyalty.

By studying literature on loyalty programs and its effect on customer churn, 14 explanatory variables have been found which possibly affect customer churn probabilities. These variables can be divided into three main dimensions; customer characteristics, loyalty program participation, and purchase behavior. These factors were tested, if possible, on a balanced sample of Menzis' customer base. Using a binary logistic regression (logit) parameter estimates have been estimated which indicate the factors influencing future churn probabilities. After correcting for missing values and a sample selection bias, several models have been constructed. The logit showed that a customer's age, gender, education level, income level, mentality group, purchase behavior, a customer's lifetime at the firm, and point collection are significant drivers of churn. Variables which do not significantly contribute in predicting future churn probabilities are social class, household size, a customer's lifetime in the loyalty program, number of different collecting variations, and point redemption (reward level). Despite the fact that a logit provided interesting parameter estimates, a classification tree (or decision tree) proved to be the most powerful model in predicting future churn probabilities.

4 SamenGezond.

Based on these findings, Menzis is recommended to focus their future customer loyalty activities mainly on the significant drivers of churn. Also the insignificant drivers of churn show insightful information because these factors may lead to ineffective customer loyalty activities. For future activities it is wise to trigger new customers to participate in the loyalty program, since they are more likely to churn and loyalty participation lowers churn probabilities.

Acknowledgements

Firstly, I would like to thank my main supervisor Tammo Bijmolt for the help, support, and feedback during the process of writing my master´s thesis. I would also like to thank Menzis for the opportunity for me to write my master´s thesis and for the help provided throughout the process. In particular I want to thank my supervisor within Menzis, Bert-Jan. Steenbergen, for the support, time, and effort invested in me while guiding me through the process. Finally I would like to send a special 'thank you' to my family, friends, and especially my fellow students for the guidance, and the unconditional support.

5

List of tables

Table 1 General mentality groups ... 14

Table 2 Healthcare specific mentality groups ... 15

Table 3 Variables used in this study ... 21

Table 4 Description of strata ... 22

Table 5 Churn divided per customer activity ... 24

Table 6 Sample size: Churn vs. non-churn ... 26

Table 7 Parameter estimates main-effects model ... 30

Table 8 Significant odd changes to churn between mentality groups ... 32

Table 9 Interaction effects for main effects model ... 34

Table 10 Key findings of hypotheses ... 39

Table 11 Variables overview ... 46

Table 12 Descriptive statistics active participants ... 47

Table 13 Descriptive statistics passive participants ... 47

Table 14 Descriptive statistics non-participants ... 47

Table 15 Descriptive statistics non-churners (churn = 0) ... 48

Table 16 Descriptives statistics churners (churn =1) ... 48

Table 17 Odd ratios between education levels ... 49

Table 18 Odd ratios between income levels ... 49

Table 19 Odd ratios between mentality groups ... 49

Table 20 Odd ratios between points collected... 50

Table 21 Odd ratios between days since last login ... 50

Table 22 Summary validation metrics ... 51

Table 23 Significant factors predicting churn ... 51

List of figures

Figure 1 Theoretical framework ... 10

Figure 2 Age dependent interactions ... 35

Figure 3 Income dependent interactions ... 35

Figure 4 Top Decile Lift curve: Logit variations ... 38

Figure 5 Top Decile Lift: Logit variations ... 38

Figure 6 Top decile lift curve: logit vs. tree ... 38

6

Table of contents

Management summary ... 3 1. Introduction ... 8 2. Theoretical framework ... 10 2.1 Customer loyalty ... 10 2.2 Customer characteristics ... 11 2.2.1 Age ... 11 2.2.2 Gender ... 11 2.2.3 Education ... 12 2.2.4 Income ... 12 2.2.5 Social class ... 13 2.2.6 Household size ... 13 2.2.7 Mentality groups... 14 2.2.8 Purchase behavior... 15 2.3 Loyalty program ... 16

2.3.1 Loyalty program participation ... 16

2.3.2 Point collection ... 17 2.3.3 Engagement ... 18 2.3.4 Lifetime ... 18 2.4 Rewards ... 19 2.4.1 Level of redemption ... 19 2.4.2 Point redemption ... 20 3. Methodology ... 21 3.1 Menzis ... 21 3.2 Data description ... 21 3.2.1 Descriptive statistics ... 22 3.2.2 Interaction effects ... 24 3.2.3 Missing values ... 25 3.2.4 Recoding variables ... 25

3.2.5 Sample selection bias ... 26

3.3 Research design ... 27

3.4 Research method ... 27

4. Model specification ... 29

5. Results ... 30

5.1 Main effects model ... 30

5.2 Main drivers of churn ... 31

5.2.1 Customer characteristics ... 31

5.2.2 Purchase behavior... 32

5.2.3 Loyalty program participation ... 33

7

5.4 Validation ... 35

5.4.1 Main effects model ... 36

5.4.2 Significant main effects model ... 36

5.4.3 Interactions model ... 36

5.4.4 Sample selection bias correction ... 37

5.4.5 Predictive validity measures ... 37

6. Discussion ... 39

6.1 Conclusions ... 40

6.1.1 Customer characteristics ... 40

6.1.2 Loyalty program participation ... 41

6.1.3 Purchase behavior... 42

6.1.4 Ineffective drivers of churn ... 42

6.2 Limitations and future research ... 43

6.3 Managerial implications ... 44

Appendices ... 46

Appendix 1: Variables used in this study ... 46

Appendix 2: Descriptive statistics per stratum ... 47

Appendix 3: Descriptive statistics churners vs. non-churners ... 48

Appendix 4: Significant differences among categories ... 49

Appendix 4a: Odd ratios for education ... 49

Appendix 4b: Odd ratios for income ... 49

Appendix 4c: Odd ratios for healthcare specific mentality groups... 49

Appendix 4d: Odd ratios for points collected ... 50

Appendix 4e: Odd ratios for last login into SamenGezond ... 50

Appendix 5: Summary validation metrics of different models ... 51

Appendix 6: Key drivers of churn ... 51

8

1. Introduction

The healthcare industry faces a big challenge in dealing with churning customers. With that, insurance companies can no longer rely on a steady customer base (Günther et al., 2014). Since 2006 the Netherlands has experienced new legislations for health insurers, called the Health Insurance Act (HIA). This HIA obliges Dutch citizens to choose a health insurer. Consumers can choose their own health insurer, as long as it acts conform a pre-determined standard (The Ministry of Health, Welfare and Sport, 2012). Dutch citizens can switch their health insurer only once a year. This in contrast to Norway, where consumers can switch their health insurance at any given moment in time (Günther et al., 2014). Also in the US the healthcare system is changing. Since 2010 every US citizen is obliged to have a health insurance. The Affordable Care Act (ACA, also known as 'Obamacare') increases competition among insurance companies, and increases cost awareness of consumers within the healthcare industry (Jones and Greer, 2013). Despite health insurance being mandatory in several countries, consumers in these countries have the possibility to switch their insurer at the end of each year. Since the healthcare industry has become a free market, consumers have become more conscious of their own healthcare policy (Stempniak, 2014). This has led to an increase in health insurers competing for the same consumers. Besides the increased industry competition, consumers also become more conscious. The increasing price consciousness may be due to the fact that consumers have to decide whether or not they want to insure for extras within their healthcare policy, like travel insurance or physiotherapy. These trends may explain the high churn rates of health insurers.

Consumers now face a tradeoff between healthcare coverage and up-front costs. Health insurers apply a cost-sharing strategy where the costs been made will be covered by all customers. Resulting in increasing prices for healthcare policies, also for those who do not consume a lot of healthcare (Stempniak, 2014; Mulvany, 2015; Gericke, 2009). In the Netherlands alone the total healthcare expenditures are expected to increase from 10% of GDP in 2010, to 24% of GDP in 2040 due to increasing healthcare consumption and the aging society (CPB, 2011). This shows that healthcare consumption can be decreased by aiming for a behavioral change.

9 value delivered to consumers (Bijmolt et al., 2010; Breugelmans et al., 2015; Astuti, 2014).

In the Netherlands alone 7.8 billion Euros is spent on health issues related to behavior, such as: obesity, smoking, or drinking (The Ministry of Health, Welfare and Sport, 2011). Changing the lifestyle of consumers can therefore be quite effective in decreasing the costs for consumers via the cost-sharing principle. Encouraging consumers to live a healthier lifestyle can decrease long-term healthcare consumption and so decrease the total healthcare costs. In order to stimulate consumers to change their current lifestyle, consumers might want to be rewarded for this effort.

In loyalty programs consumers are often rewarded for their behavior. However, most loyalty programs are built on purchase frequency: the more you buy, the more you get. Many studies show how loyalty programs affect purchasing behavior in industries like retail or travel (e.g.: Zhang and Breugelmans, 2012; Dorotic et al., 2012; Bijmolt et al., 2010; Eason et al., 2015; Vorhees et al., 2015; JiungYee Lee, 2014; Wansink, 2003). Within the healthcare industry increasing sales might not be the most obvious goal. After all, one may assume that healthcare is here to make people healthier. For firms in the healthcare industry it might be more appropriate to change consumers' lives compared to e.g. retail companies.

To study whether loyalty programs, which encourage a healthy lifestyle, can reduce customer churn, data from Menzis, a major Dutch health insurer, is used. The contribution of this study to existing literature consists out of the investigation of the effect of loyalty programs aimed at changing one's behavior. A loyalty program will be studied, which stimulates a healthy lifestyle so that long-term healthcare consumption decreases. The purpose of this study is to provide a model which predicts churn for Menzis, aimed at their loyalty program which adds value for the customer by changing customer's lifestyle so that churn probabilities decrease (instead of the more common goal of loyalty programs: increasing sales). The main research question is to identify those factors which affect customer churn and the role of the loyalty program in this process.

10

2. Theoretical framework

In order to derive the hypotheses for this study existing literature is studied. First the dependent variable will be shortly introduced. This study aims to predict customer loyalty. The concepts of customer loyalty, retention, and churn are all very much related. In this study customer loyalty is defined as those customers who stay at the firm after the yearly churn opportunity has expired (in other words, those customers who do not churn). This can best be illustrated by the following expression: . This way customer loyalty can be estimated by predicting the probabilities that consumers are going to churn. In this chapter several possible concepts which influence customer loyalty will be introduced. Figure 1 shows a visual representation of the theoretical framework of this study. Based on literature, expected outcomes will form the foundation of the hypotheses.

Figure 1 Theoretical framework

2.1 Customer loyalty

11 customer loyalty is limited to the actual behavior of customers (behavioral loyalty). Loyal customers will therefore be defined as: "customers who stay with the firm after the yearly opportunity to churn has expired" (in other words; the customers who do not churn).

Loyal customers are crucial for any firm. The financial risk of customers leaving your company (churn) is enormous. With the increasing realization that attracting new customers results in higher expenses compared to retaining your current customers, the importance of customer loyalty is evident (Too et al., 2001).

2.2 Customer characteristics

There is no such thing as 'the average customer'. Therefore firms who want to tailor their offerings to individual needs often work with customer segments. Each customer segment has its own characteristics, which separates them from other segments. Based on these differences it is to be expected that the effect of customer characteristics determine whether a consumer is likely to participate in a loyalty program. One way to distinct customers from each other is on the demographic level. In this study the effects of some demographics will be studied and explained in the following paragraphs.

2.2.1 Age

Within the Dutch healthcare industry there is a trend where the majority of the annual churners consists out of young consumers (KlantenMonitor Zorgverzekeringen, 2013). Younger people tend to be more price conscious and are more likely to be looking for cheaper alternatives each year (Wong, 2011). The reason behind this trend is that this customer group does not consume a lot of healthcare. A study by Wong (2011) showed that a customer's desire to churn decreases when the customer is getting older. In a purchase decision process older people tend to base their purchasing decisions more on previous experiences such as satisfaction, whereas younger people rely more on information provided by sales personnel (Homburg and Giering, 2001).

This information about age differences on churn behavior suggests a negative effect of age of the consumer on churn, so a positive relation between age and customer loyalty.

 H1: Age is negatively related to churn.

2.2.2 Gender

12 brand. This shows that women are a more lucrative target group when it comes to attracting new customers.

Women process more information when deciding where to purchase compared to men. Where men decide where to purchase based on previous experiences (e.g. satisfaction level) women decide based on more information (e.g. interaction with people during the purchasing process) (Homburg and Giering, 2001; Mitchell and Walsh, 2004). Customers who are willing to seek for more information to base their purchase decision on, are more likely to switch their current supplier (Homburg and Giering, 2001). When customers are open for new information, customers also may come across new suppliers which may offer better solutions compared to their current supplier. Whereas men remain loyal to the firm when the positive satisfaction level is held, women do not solely rely on the satisfaction level to remain loyal (Homburg and Giering, 2001). Women are more involved in purchasing activities, therefore it may be expected that women are pickier when choosing the right offer compared to men. This infers that women are more likely to switch their current provider when an alternative solution is offered compared to men. Based on these results the hypothesis is formed that men are less likely to churn.

 H2: Men are less likely to churn, compared to women.

2.2.3 Education

There seems to be little evidence for education in relation to customer loyalty. However, literature shows that higher educated people engage more in information gathering, information processing and use more information for decision making. Lower educated people rely on less information for decision making (Evanschitzky and Wunderlich, 2006). Furthermore, higher educated people are more comfortable with processing new information in the decision making process (Homburg and Giering, 2001). Following this information infers that higher educated people are more likely to seek for alternative offerings. Therefore it is to be expected that higher educated people are more likely to churn. Literature also shows that higher educated people are more willing to participate in a loyalty program compared to lower educated people (Dorotic et al., 2012). With higher educated people being pickier about firms, a positive relation is expected between education level and churn.

 H3: Education level is positively related to churn.

2.2.4 Income

13 seek for alternatives would infer that consumers with a lower income are more loyal to a firm compared to consumers with a higher income.

 H4: Income level is positively related to churn.

2.2.5 Social class

Little research is known about the influence of social classes on customer loyalty. However, a study by Martineau (1958) shows that shopping behavior can be strongly related to purchasing behavior. As the study from the late 50's shows, certain social classes can gain great social status based on the places where they purchase. This study is however aimed at purchases in general. It is questionable whether this could also be the case for insurance policies. The concept of social status appears to be also relevant in more recent studies. Shopping motivations can be seen as an expression of social and political relations between households (Crockett and Wallendorf, 2004). Both studies show that differences in social class can make a difference in buying behavior. Based on elements like status and preferred behavior within groups, choices can be influenced. This triggered the formulation of the following hypothesis where a higher social class results in a lower churn probability:

 H5: Social classes are negatively related to churn.

2.2.6 Household size

To see whether household size has an impact on churn behavior, evidence has been found in the study of Yang (2014). Yang studied the Danish retail electricity market and shows that the market is actually quite comparable to the Dutch healthcare industry because of two reasons: Firstly, the market is highly competitive and secondly, switching is a simple process which can be done by phone or internet. This way there are low switching costs for consumers wanting to switch their current service provider.

Yang (2014) even states that consumers only switch whenever they can save costs. This argument is also supported by Haisley and Loewenstein (2011), they state that firms become less attractive when the perceived benefits of that firm decrease.

Relating this cost consciousness to household size, we see that larger households are in general more price conscious (Glynn and Chen, 2009). Glynn and Chen (2009) have found that larger households are more price conscious and are therefore more prone to purchase private label brands. Extending this way of thought to this study in the Dutch healthcare industry, we can say that larger households are more likely to churn whenever they find a cheaper offer. Together with the fact that the Dutch healthcare industry is highly competitive, and competitive prices are likely to occur, the next hypothesis is formulated:

14

2.2.7 Mentality groups

For this study the segmentation based on the mentality groups (Motivaction, n.d.) will be leading. With eight pre-determined groups, consumers can be divided into segments based on two dimensions (status and social standards). The mentality groups have proven to be a rich tool for explaining consumer behavior by several demographic elements. These groups go further than the earlier mentioned demographic variables because the mentality groups incorporate more dimensions than the ones available for this study. By including the mentality groups as a unique variable in the model, the model's predictive power will increase.

The eight segments Motivaction (n.d.) has defined are; Traditionals, Modern mainstream, New conservatives, Post-materialists, Postmodern hedonists, Convenience oriented, Social climbers, and Cosmopolitans. The general mentality groups are applicable for any industry (table 1).

Table 1 General mentality groups

General mentality groups Description

Traditionals The moralistic, dutiful, status-quo oriented citizens who hold on to traditions and material assets.

Modern mainstream The conformist, sensitive to status who seeks for the balance between tradition and modern values like consuming and enjoying and tries to adapt to the environment.

New conservatives The conservative liberal upper society who loves technological innovation and protests against social and cultural innovation.

Post-materialists Idealist, critical at the society and wants to evolve itself, standing-up for social injustice and the environment.

Postmodern hedonists The pioneer of the experience culture who likes to experiment and enjoys the collision with moral and social values.

Convenience oriented The impulsive and passive consumer who strives for a carefree, happy and comfortable life.

Social climbers The career oriented individualist with a outspoken fascination for social status, new technologies, risk and excitement.

Cosmopolitans Critical cosmopolitan who integrate post-modern values like evolve and experience with modern values like success, materialism and pleasure.

15 Table 2 Healthcare specific mentality groups

Healthcare mentality groups Description (1) Consumption oriented healthcare

client

The conformist, status-conscious citizenry that seeks a balance between tradition and modernity values consume and enjoy.

(2) Idiosyncratic healthcare client Pioneers with the experience culture, in which experimentation and breaking with moral and social conventions are leading.

(3) Convenience oriented healthcare client

The impulsive and passive consumer who strives first and foremost to a carefree, enjoyable and comfortable life.

(4) Quality oriented healthcare client The open and critical global citizens, who integrate postmodern values and experiences with modern values such as social success, materialism and enjoy. (5) Luxurious oriented healthcare

client

The liberal-conservative upper class, who wants to make room for technological development, but opposes social and cultural innovation. (6) Social critical healthcare client The socially critical idealists who want to develop themselves, oppose social

injustice and stand up for the environment.

(7) Pragmatic healthcare client The career-minded individualists with a pronounced fascination with social status, new technology, risk and pressure.

(8) Compliant healthcare client The moralistic, dutiful and on the status quo-oriented citizenry who clings to traditions and material possessions.

Each mentality group is known for their own behaviors, motives, and preferences. Based on these individual differences, it is to be expected that the different mentality groups have different effects on customer loyalty.



H7: Churn differs between the mentality groups.

2.2.8 Purchase behavior

With purchase behavior we aim at the purchase frequency, and purchase amount. In this study, 'purchase frequency' refers to the number of times a consumer changes their healthcare policy each year (e.g. add or remove additional policies, increase or decrease legally own risk). 'Purchase amount' refers to the number of policies a customer has with Menzis (e.g. only basic insurance, or basic insurance plus multiple additional policies). Note that consumers can only have one health insurer at a time and can only churn once a year, but in some occasions can change their additional policies throughout the year (depending on the health insurance company). The focal firm of this study, Menzis, does not allow changes in additional policies throughout the year (except when your child reaches an certain age) (Menzis, n.d.). Purchase behavior in this study will therefore be limited to the number of healthcare policies a consumer has with Menzis.

Literature shows that purchase behavior can have significant effects on customer loyalty via trust. Consumers who purchase at a firm show their trust in that firm (Collado Agudo et al., 2012). When consumers show their trust in a firm by purchasing there, they are becoming loyal to the firm. However, in this study we are mainly interested in the purchase behavior aspects that are directly linked to churn by means of behavior (e.g. purchase frequency and purchase amount).

16 their healthcare policy while remaining to be a customer at one specific firm, their likelihood to churn would decrease. Customers who change their healthcare policy during the observation period basically increase their purchase frequency and are therefore more loyal. It did not matter whether they changed their healthcare policy once, or multiple times. They are significantly more loyal to the firm. Based on the information mentioned above it is to be expected that the churn probability of customers is affected by their purchasing behavior.

 H8: Purchase behavior influences customer loyalty where purchase amount has a negative effect on churn.

2.3 Loyalty program

A better understanding of the systematic factors which influence loyalty program performances is important because it can help managers to understand whether a loyalty program is an appropriate instrument in a certain context and can help improving the effectiveness of existing loyalty programs (Liu and Yang, 2009). Following the definition of Leenheer et al. (2007), loyalty programs can be defined as "an integrated system of marketing actions that aim to make member customers more loyal". Another much adopted definition of loyalty programs is: "Continuity incentive programs offered by a retailer to reward customers and encourage people to purchase products on a continuing basis or over time." (The American Marketing Association's dictionary, n.d.). This definition is sufficient, except for the fact that it only focuses on the retailers and suggests that people get rewarded based on purchase behavior (Bijmolt, et al., 2010). Loyalty programs can be adopted in a wide range of industries, like the financial services, travel or retail industry. Common loyalty programs reward consumers for frequent purchases, the so called frequency reward programs. In this study the loyalty program focuses on a change in lifestyle, but in the end still provides customers with saving some sort of currency. These savings can be redeemed for later rewards (Liu and Yang, 2009).

Another major difference between the loyalty program being studied and frequency reward programs is that a consumer can only have one health insurance, which excludes polygamous loyalty. Polygamous loyalty is quite common in industries like retail or travel where consumers purchase and repurchase different brands simultaneously, but it does not apply to this study where health insurances are studied (Ferguson and Hlavinka, 2007; Dawes, 2009; De Wulf et al., 2003; Lucas et al., 2005).

2.3.1 Loyalty program participation

17 programs are known for their benefits to the firm. However, there remain ambiguities about the benefits for the customers because mostly the firm would benefit from the loyalty program (Mimouni-Chaabane and Volle, 2010). Despite the high level of interest in academic literature and the increasing popularity of loyalty programs in a variety of industries during the past decade (such as retail, financial services and travel) (Nielsen’s Global Survey of Loyalty Sentiment, 2013), empirical studies provide multiple viewpoints on loyalty program effectiveness in cementing customer loyalty (McCall & Vorhees, 2010; Leenheer et al., 2007; Dorotic et al., 2012). Marketing researchers throughout the world have found positive effects of loyalty programs on customer loyalty (e.g.: Zhang and Breugelmans, 2012 Dorotic et al., 2012; Bijmolt et al., 2010; Eason et al., 2015), as well as results that show that loyalty programs may not considerably affect customer loyalty (e.g.: Mägi, 2003; Meyer-Waarden and Benavent, 2006; Nunes and Drèze, 2006). This confusion about the effect of loyalty programs illustrates why loyalty programs are still extensively studied. Part of this confusion may be based on a lack of evidence for the drivers of loyalty program effectiveness. Because most literature states that consumers who participating in loyalty programs will (at least to some degree) increase loyalty, the following hypothesis can be formulated:

 H9:Loyalty program participation is negatively related to churn.

2.3.2 Point collection

Point collection can be done in various ways. A common way to collect points within a loyalty program is based on frequent purchases. In this study however, points can be collected when a customer adapts to a healthy lifestyle. Either way, points are given in return for a change in behavior. Subsequently, the points can be collected and redeemed for products or services on a later moment in time. Consumers may want to save their points in order to be able to purchase a more luxurious product to reward themselves (or others). This point collection aims to let customers feel more connected to the firm. Collecting points can only be done whenever they stay loyal to that specific firm.

18 firm. This connection of the point-pressure mechanism, together with the illusion of advantage, results in more loyal customers and the formulation of the next hypothesis:

 H10: Point collection (h10a) and the level of point collection (h10b) are both negatively related to churn.

2.3.3 Engagement

Engagement can be defined as consumers' positive, fulfilling, brand-use-related state of mind (Dwivedi, 2015). Highly engaged consumers are connected more and have stronger connections with a firm. In this study, engagement does not account for the level of engagement with the brand, but with the loyalty program. Because loyalty programs are studied instead of brand or firm associations, highly engaged customers can be defined as: "those customers who use the loyalty program very often". This infers that highly engaged customers are more likely to become loyal to the firm. One could say that a consumer's motivation to participate in a loyalty program is quite self-centered. Their decision whether or not to participate in a loyalty program is based on a reflection of the potential benefits loyalty program (hedonic, symbolic or utilitarian) relative to the perceived input (cost, risk or effort) of this enrollment (De Wulf et al., 2003; Bijmolt et al., 2010; Dorotic et al., 2012). This self-oriented motivation to participate in a loyalty program can have major effects on the effectiveness of the loyalty program. Highly engaged customers are assumed to have the highest self-oriented motivation and can obtain the most benefits. Leenheer et al. (2007) refer in their study to the self-selection or endogeneity problem. This self-selection or endogeneity problem infers that customers who are already engaged to the firm may be the ones who are most likely to participate in any loyalty program. In other words, those customers who are loyal and are frequent purchasers are the customers who can obtain the most benefits without adjusting their behavior a lot (Bijmolt et al., 2010). This could lead to an overestimation of the effect of a loyalty program on customer loyalty and should therefore be taken into account when studying loyalty program effectiveness.

 H11: Engagement level is negatively related to churn.

2.3.4 Lifetime

The lifetime can be defined as the time someone is a customer at the same firm, or in other words: relationship duration. For health insurers this relationship duration can be of great importance. The so called cherry pickers in the insurance industry always seek for the best alternative. Chances are that customers with the shortest relationship length are more likely to churn, in comparison to customers who have been loyal for many years.

19 length, the lower the customer changes his or her behavior. In other words, the longer someone is a customer at a firm, the more likely he or she is going to stay with that firm.

 H12: Customer lifetime (H12a) and duration of loyalty program participation (H12b) are both negatively related to churn.

2.4 Rewards

Rewards can make or break any loyalty program. The main reason of many customers to participate in loyalty programs is not to become loyal to the firm, but to obtain the rewards provided with the loyalty program. Literature also states that customers, in general, rather do not want a long-term relationship with a firm (Liu, 2007). This is in contrast to the fact that, from the firm's point of view, loyalty programs are generally introduced to establish long-term relationships with customers. In fact, customers are focused on rewards (immediate or postponed). Immediate rewards are aimed at attracting new customers on the short-term and provide customers with immediate rewards when purchasing a product. In order to reach long-term customer loyalty, loyalty programs can offer postponed rewards (where you first collect points and save these points, in order to spend them later in time) (Park et al., 2013; Zhang et al., 2000; Hsee, et al., 2003).

2.4.1 Level of redemption

In this study 'reward' is interpreted similar to redemption. The points that are redeemed represent the reward a customer obtains. The benefits a consumer can gain by participating in a loyalty program are important reasons to participate in any loyalty program (e.g. Park et al. 2013). When the benefits outperform the perceived costs of participating in a loyalty program, the likelihood of entering that loyalty program increases. This effect is consistent with a study by Kim et al. (2012) where they mention that whenever a consumer perceives or expects great benefits of a loyalty program, they are more willing to participate in this particular loyalty program. This leads to greater relationship proneness by consumers, which is positively related to the perceived benefits of a loyalty program and negatively to the resistance to change (Kim et al., 2012). Also, when customers receive a reward their expectations of subsequent rewards go up. Customers expect the subsequent reward to have similar or even better perceived benefits. Customers might react negatively when the subsequent rewards have lower perceived values than they might have expected (Haisley and Loewenstein, 2011).

Based on the illusion of advantage (see section 2.3.2 Point collection) and increasing relationship proneness when the reward increases the following hypothesis can be stated:

 H13a: The redemption level is negatively related to churn.

20

2.4.2 Point redemption

In addition to the level of point redemption (see previous section 2.4.1 Level of redemption) this study also aims to test the effect of point redemption as such on churn. Whether or not customer redeem points can show different results compared to the level of point redemption.

Point redemption refers to the points customers can spend which they have collected via loyalty programs. Once they have collected enough points they can redeem or spend the points to obtain incentives, rewards, or discount. Consumers who redeem their points at a certain moment in time, become more loyal to the firm due to changing purchase behaviors. The process of point redemption reinforces the customer relationship over time (Smith and Sparks, 2009).

In a study by Kumar and Shah (2004) they mention that the key goal of rewarding consumer behavior is to establish reciprocity. The authors define reciprocity as; 'customers evoking obligation towards companies who treat them well or provide value'. By providing customers with some incentive, reward or discount consumers feel obliged to react in a way which may lead to more business which in turn can lead to more rewards, and so on. This vicious circle builds on long-term customer relationships which increases customer loyalty (Taylor and Neslin, 2005; Nunes and Drèze, 2011). This vicious circle is also known as the rewarded-behavior mechanism, where participants act differently after they are rewarded in such a way that reinforces their relationship with the firm (Bijmolt et al. 2010). This information has led to the next and final hypotheses:

21

3. Methodology

This study aims to highlight the factors which predict future churn probabilities within the Dutch healthcare industry. For this study data from Menzis, a major Dutch health insurer, is used to study the effect of the loyalty program SamenGezond on churn probabilities. Due to privacy issues, not all variables that were available could also be used in this study. The variables that were used are shown in table 3.

Table 3 Variables used in this study

Independent variables Description Notation

Age Age of customer. Range 1-103

Gender Gender of customer. Binary 0=male,

1=female Points collected Sum of number of points collected for the SamenGezond program at the

end of the observation period.

Categorical 1-4

Point collection variations

Number of different ways to collect points for the SamenGezond program.

Categorical 1-3

Points redeemed Sum of number of points redeemed for the SamenGezond program at the end of the observation period.

Categorical 1-4

Days until most recent login

Amount of days it has been since the last time the customer has logged into SamenGezond.

Categorical 1-3

Household social class The social class of the customer, on a household level. Categorical 1-3 Household education

level

The level of education of the customer, on a household level. Categorical 1-3

Household income level The income level of the customer, on a household level. Categorical 1-3 Household mentality

group

The healthcare specific mentality group (Mentality, n.d.) accounting for a specific customer, on household level.

Categorical 1-8

Household size The amount of people in the in the household of the customer. Numeric Lifetime Menzis The moment in time when the customer has subscribed for an insurance

at Menzis for the first time.

Numeric / scale

Lifetime SamenGezond The date a customer subscribed for the SamenGezond program. Numeric / scale Dental care Whether or not a customer has an additional healthcare policy (dental). Binary 0=no, 1=yes Additional policy Whether or not a customer has an additional healthcare policy (general). Binary 0=no, 1=yes Churn Dependent variable, binary variable which indicates whether or not a

customer stays at the focal firm after the observation period.

Binary 0=no churn, 1=churn

3.1 Menzis

With over two million insured consumers, Menzis is a leading health insurer in the Netherlands. Menzis is socially involved and wants to trigger consumers to adopt a healthier lifestyle. With Menzis´ loyalty program, called Samengezond, they stimulate a healthier lifestyle. Via an mobile application and website customers can collect points by showing they are living healthy (e.g. show runs via running apps like Strava or RunKeeper, eat healthy, or when customers are volunteering in the society). The collected points can be redeemed on a later moment in time, in return for discount on their healthcare policy or for small gifts from the corresponding webshop. Samengezond has evolved to a large platform with over 500.000 participants in 2015.

3.2 Data description

22 For this study a stratified sample out of the customer base of Menzis has first been drawn (n = 57.000). A stratified sample is a sampling method which first divides the target population is several mutually exclusive segments (the so called strata), with respect to one or more common characteristics (Blattberg et al., 2008). In this study a segmentation has been made based on loyalty program participation (table 4). A stratified sample was first preferred over a random sample because stratified sampling guarantees more representative samples with respect to the target population. Stratified sampling increases statistical efficiency when the strata are homogeneous within groups (Blattberg et al., 2008). However, after sampling the stratified sample showed some difficulties as explained below.

Table 4 Description of strata

Subgroup of the population Description Original sample size Stratum size (balanced) Active participants of

Samengezond

Those who collected or redeemed at least one point within the observation period.

n = 20,279 n = 19,000

Passive participants of Samengezond

Those who participated in the program, but did not collect or redeem at least one point within the observation period.

n = 19,782 n = 19,000

Non-participants of Samengezond

Those who did not enroll for the Samengezond program.

n = 23,723 n = 19,000

3.2.1 Descriptive statistics

The strata active participants, passive participants and non-participants have 20.279, 19.782, 23.723 data points respectively and all have been collected from a sample of the original customer base of Menzis. Because churn behavior differs between the strata (F(2, 11175) = 121,361, p = 0.00), the descriptive statistics are now specified per stratum. A summary of all descriptive statistics, per stratum, is specified in appendix 2 and a more detailed description of the used variables is specified in section 4 Model specification.

Active participants

The active participants are on average 47 years old. This stratum is called 'active participants' because they have collected or redeemed at least one point within the observation period. The active participants are equally divided by gender (53.7% male, 46.3% female). On average active participants collected 179 points, redeemed 62 points, and have used two point collection variations. Of those participants who collected points, 4.3% also redeemed an amount of these points.

Passive participants

23 The same holds true for the 430 customers who have logged into the SamenGezond program. This can be explained because these customers have been active SamenGezond participants in the past (before the observation period). Even though this group of 434 customers have shown sort of activity, they are processed as passive-participants because participants are only considered being active when they collect or redeem points within the six months of the observation period.

Non participants

The non-participants stratum is meant to validate the model and as a reference category to test whether the SamenGezond program as such has an impact on churn probabilities. The non-participants are on average 47 years old and are equally divided by gender (male 49.6%, female 50.4%). As mentioned above, non-participants do not have any observed values for SamenGezond usage specific variables like point collection, last login, or point redemption because they have not participated during the observation period.

Churners vs. non-churners

In this section the descriptive statistics of the balanced sample being used in this study (churners vs. non-churners) are provided. The average age of the balanced sample is 39 years, and the sample is equally divided on gender (male 49.3%, female 50.7%). Despite the non-churners are drawn at random from the original sample, they are divided quite equally. The sample is divided into active participants (34.8%), passive participants (33.8%) and non-participants (31.5%). A more detailed specification of the descriptives, divided per segment (churn vs. non-churn), is provided in appendix 3.

24 Table 5 Churn divided per customer activity

n # Churn % of total churn

% churn relative to category Segments Non-participants 23723 1380 24,70% 5,82% Passive participants 19782 2050 36,69% 10,36% Active participants 20279 2158 38,62% 10,64% Total 63784 5588 100% 8,76% Point collection Collectors 20263 2156 38,58% 10,64% 1-500 19352 1997 35,74% 10,32% 501-1000 715 135 2,42% 18,88% 1001-1500 112 14 0,25% 12,50% 1501-2000 28 5 0,09% 17,86% 2001 > 56 5 0,09% 8,93% 0 (Non-collectors) 43521 3432 61,42% 7,89% Total 63784 5588 100% 8,76% Point redemption Redeemers 890 168 3,01% 18,88% 1-500 245 56 1,00% 22,86% 501-1000 166 26 0,47% 15,66% 1001-1500 302 55 0,98% 18,21% 1501-2000 59 14 0,25% 23,73% 2001 > 118 17 0,30% 14,41% 0 (Non-redeemers) 62894 5420 96,99% 8,62% Total 63784 5588 100% 8,76%

3.2.2 Interaction effects

25

3.2.3 Missing values

The stratified sample showed many missing values, which needed to be adjusted for. To test whether the missing data points can be replaced with expected values Little's MCAR (Missing Completely At Random) was conducted to test if the missing values were missing at random. Little´s MCAR test showed that the missing values are missing not at random (MNAR), p > 0.05. Therefore the missing values could not be replaced with predicted values, maximum likelihood imputation or list wise deleting (Myers, 2011). Instead the missing values represented some sort of non-response (e.g. customers who did not participate in the SamenGezond program automatically also did not collect points, resulting in a missing value for that specific variable). By transforming all missing values (if this meant 'no', 'zero', or 'none') into a '0', the missing values have a value which can be interpreted (namely that this variable was not entered, unknown, or zero). When this transformation was skipped, all missing values were automatically excluded from the analysis. By transforming the data such that each individual data point has a meaningful value, (e.g. ´0´ or ´1´, instead of ´missing value´ or ´yes´), the model performance improved greatly because now a much larger number of cases can be included in the analysis. After correcting for the missing values the predictions were still biased by the rare events case (see section 3.2.5 Sample selection bias).

3.2.4 Recoding variables

The original dataset consisted out of multiple categorical variables. These categorical variables varied between 4 and 8 categories. In order to improve the interpretation of the model among the different categories of the independent variables, a new three-way distribution has been chosen if possible (1 = low, 2 = medium, 3 = high). This three-way distribution was chosen because this improved the relative distribution based on frequencies (the groups become more equal in terms of number of customers in each group). Variables like income, social class, education level, level of point redemption, level of point collection, days until last login into SamenGezond, and number of different point collection variations were recoded into the new three-class distribution. This improved interpretability of the parameter estimates, which enforces the model to be more parsimonious (simple, but still complete).

26 A customer's lifetime in the SamenGezond program was recoded into months. Customers can enter SamenGezond at any moment in time, therefore months was more applicable than years. Despite the effect size improved the variable Lifetime in SamenGezond remained insignificant (p value > .05), as shown in section 5. Results. This insignificance holds when recoding into one unit = 30 days, one unit = 90 days, and one unit = 365 days.

3.2.5 Sample selection bias

Based on the data it can be seen that churning is a quite rare event occurring in the data set. Of the total customers available in this study only 8.8% did churn after the observation period. Since the occurring event is rather low (churn = 1), the probabilities to churn are low. These so called rare events data may lead to an under-estimation of customer churn probabilities when conducting a binary logistic regression ( < .5 → P = 0). According to King and Zeng (2001) predicted churn probabilities become more biased when the occurrence of the event (churn = 1) is very low. The cases which are showing the event of churn might not be sufficient to fully cover the tail of the logistic distribution (Blattberg et al., 2008). This bias under-estimates the probability of customers to churn, resulting in an estimation which is too pessimistic (less than 1% predicted churners with the stratified sampling method).

Despite the 8.8% of annual churners reflects a real-life situation, the vast majority of non-churners dominate the statistical analysis, which may lead to decreased predictive accuracy (Lemmons and Croux, 2006). A method to correct for this sample selection bias, as proposed by Donkers et al. (2003), is by oversampling the outnumbered group (the churners in this case). With oversampling the outnumbered group, the relative share of the outnumbered group within the sample increases significantly. This leads to obtaining a new sample with two groups equal in size (churners vs. non-churners), as shown in table 6. A balanced sample is preferred in this study over a stratified sample with respect to the predictive power of the model. Because the balanced sampling method represents a disproportionate amount of the population, the sample will be corrected to account for the actual population size. Literature proposes a weighted correction method for the balanced sampling method that may help in constructing a better churn prediction model (Donkers et al., 2003; Lemmons and Croux, 2006; King and Zeng, 2001). This weighted scale is further explained in section 3.4: Research method of this study. Oversampling the churners (Y = 1) in this case only affects the constant of the model, parameter estimates will remain the same (King and Zeng, 2001). This supports the reasoning for choosing a new balanced sample.

Table 6 Sample size: Churn vs. non-churn

Subgroup of the population Description Original sample size Sample size (balanced) Churners Those customers who did churn after

the observation period ends.

n = 5,588 n = 5,588

Non-churners Those customers who did not churn after the observation period ends.

27

3.3 Research design

Based on a cross-sectional data sample from Menzis' customer base, a binary logistic regression model (logit) will be constructed. With a logit we estimate the probabilities of customers going to churn. From utilities, churn probabilities are calculated which can be translated to choice (churn: yes/no). This model provides Menzis which information on which factors affect churn probabilities among their customers. By constructing a churn prediction model, this study aims to demonstrate which elements affect customer churn within the health insurance industry. In general this study is descriptive, rather than predictive. By means of a churn prediction model the relevant factors that affect churn predictions are obtained.

3.4 Research method

In order to be able to predict customer loyalty for Menzis, a binary logistic regression is used (logit) resulting in future churn probabilities. With a logit regression customer loyalty is calculated from utilities, via probabilities to a binary choice of churn: yes/no (equation 1). Based on the advantage of interpretation, a logit is preferred over a similar approach such as a probit because the outcomes are changes in odds. Whereas a probit provides changes in z-scores.

[1]

Based on a set of parameters the preference of customer i to stay with the firm, or to churn, is based on utilities ( , see equation 2.

[2]

The probability of observing = 1 for customer i, given , is equal to the cumulative distribution function evaluated at the utilities. This can be specified as:

[3]

Transforming utilities into probabilities, by applying the logistic cdf (equation 3), results in an estimation for the latent variable; , indicating the probability that customer i will cancel all insurance policies at Menzis in the next year (Y = 1). The logit will calculate the probability that customer i is going to churn. For customer i (i = 1, ..., n) an estimated churn probability > .5 will most likely result in churn in the next year (equation 4). The unobserved variable can be interpreted as the estimated churn behavior for individual i. The relation between and the observed choice can be specified as:

[4]

28 An important criterion to interpret the parameters of the logit model is the odds ratio (equation 5).

[5]

Odd ratios show the probability of an event to occur, relative to that event not to occur. An odd ratio of two means that the probability of is two times larger than the probability of . In logit models, it is common to interpret odd ratios as a log odds ratio (equation 6), which improves interpretability of the odd ratios.

[6]

Correcting for a balanced sampling method

As mentioned before this study shows some sample selection bias, which needs to be corrected for. As proposed by literature there are basically two easy to interpret methods to correct for the sample selection bias (Lemmons and Croux, 2006; King and Zeng, 2001). The first alternative is by adding a weighted scale to the estimated churn probabilities. This study corrects for the number of churners in the balanced database by adding a weighted scale to the estimated probabilities, relative to the actual number of churners in the original dataset. The probability to churn ( ) in the stratified sample will be corrected with the weighted scale as specified in equation 7.

[7]

and;

The in this study is the overall churn rate (8.8%), where N is the total number of customers in this sample (57,000). These weights can be attached to the estimated probabilities of churners and non-churners respectively. Note that the sum of the weights in equation 7 always equals one. Adding weighted scales to the estimated probabilities is statistically a valid approach to correct for stratified sampling (Lemmons and Croux, 2006). However, because the weights assigned to churners are small, this correction cancelled the advantage of oversampling and provide similar results as the stratified sampling method. The results of the weighted correction are shown in the section 5.4.4 Sample selection bias correction.

29

4. Model specification

In the next section the starting model used for predicting churn probabilities for customers in the Dutch health insurance industry is presented (equation 8).

[8]

Where for customer i,

U The utility that customer i obtains for going to churn in the next period.

AP Binary variable whether or not customer i has an additional healthcare policy.

DP Binary variable whether or not customer i has an dental care policy.

SG Binary variable which indicates whether or not customer i is participating in the SamenGezond program.

Age Age of the customer (in years).

Gender Gender of the customer (0=male, 1=female).

LC Lifetime at the company (Menzis), one unit equals 365 days.

LSG Lifetime at the loyalty program (SamenGezond), one unit equals 30 days.

COL_bin Binary variable whether or not customer i has collected at least one point.

RED_bin Binary variable whether or not customer i has redeemed at least one point.

PERS Number of persons per household.

SOC_ Social class per household (dummy coding; low, medium, and high)

EDU_ Education level per household (dummy coding; low, medium, and high)

INC_ Income level per household (dummy coding; low, medium, and high)

COL_ Number of points collected (dummy coding; 1-50, 51-100, 101-200, 200 >)

RED_ Number of points redeemed (dummy coding; 1-500, 501-1000, 1001-1500, 1500 >) LL_ Number of points collected (dummy coding; < 1 month, 2-3 months, 3 months >)

COLVAR_ Number of different variations/alternative ways used to collect points within the observation period (dummy coding; 1/2/3 >).

CC_ Mentality group (healthcare specific pre-determined segmentation groups) (dummy coding;

Consumption oriented healthcare client/Idiosyncratic healthcare client/Convenience oriented healthcare client/Quality oriented healthcare client/Luxurious oriented healthcare client/Social critical healthcare client/Pragmatic healthcare client/Compliant healthcare client)

30

5. Results

The next section will describe the results of the model as specified in equation 7. After the parameter estimates, the main effects are described and the constructed models are validated.

5.1 Main effects model

The central model obtains the estimated parameter coefficients of the main-effects model (table 7).

Table 7 Parameter estimates main-effects model

Variable S.E. P-value Exp(B)

Customer characteristics Age in years .001 .002 .995 Gender (1) .048 .008 1.137 Social class Low .159 .709 1.061 Medium .141 .182 1.207 High * Education Low .111 .001 .699 Medium .091 .030 .821 High * Income Low .144 .460 .899 Medium .073 .019 .844 High *

Persons per household .022 .992 1.000

Care consumer (mentality group)

Consumption oriented healthcare client .110 .063 1.228 Idiosyncratic healthcare client .136 .042 1.317 Convenience oriented healthcare client .131 .143 1.113 Quality oriented healthcare client .125 .041 1.292 Luxurious oriented healthcare client .136 .891 1.019 Social critical healthcare client .124 .743 1.042 Pragmatic healthcare client .122 .038 1.288 Compliant healthcare client *

Purchase behavior Lifetime company in years (date of first policy) .002 .000 .879

Additional insurance policies (1) .058 .000 .766

Dental insurance policy (1) .050 .815 .988

Loyalty program usage Lifetime SamenGezond in months(date of first enrollment) .012 .098 1.075

31

5.2 Main drivers of churn

Interpretations of the parameter effects in a logit model is done by calculating the odds ratio: The probability that Yi = 1, divided by the probability that Yi = 0 (Leeflang et al., 2015). For example, an

odds ratio of five means that the probability that Yi = 1 is four times larger than the probability that Yi

= 0. An alternative interpretation of the odds ratio of five is that the probability that Yi = 1 is 400%

larger than the probability that Yi = 0. In this section the parameter estimates for the main effects

model are interpreted in odd ratios (exp(B)) to find the main drivers of churn.

5.2.1 Customer characteristics

The age of the customer significantly affects customer churn (p < .05). The estimated coefficient of age on churn is negative (exp(B) < 1). The odds for people to churn decreases with .5% when people age with one year. This means that older people are less likely to churn whereas younger people are more likely to churn. Given this information the data confirms hypothesis 1.

For gender the estimated effect is positive and significant (p < .05, and exp(B) > 1). The odds to churn for women are 13.7% higher compared to men. This means that men are 1.137 times less likely to churn compared to women, which confirms hypotheses 2.

The level of education per household is significant and negative for both the lower educated people as well as the medium educated people in comparison to the higher educated people (p < .05, exp(B) < 1), which confirms hypothesis three (see also appendix 4a: Odd ratios for education). The odds for lower educated people to churn are .690 times lower compared to higher educated people. The odds for medium educated people to churn are .821 times lower compared to higher educated people. The odds for lower educated people to churn are .839 times lower compared to medium educated people, and the odds for higher educated people to churn are 1.209 times higher compared to medium educated people. The odds for higher educated people to churn are 1.370 times higher compared to lower educated people.

The income level per household is partially significant in relation to churn. Low-income households do not churn significantly more, compared to the households who earn a medium or higher income (see also appendix 4b: Odd ratios for income). For households with a medium income however, the odds to churn are .844 times lower compared to those households who earn a high income (p < .05, exp(B) < 1). This means that hypothesis 4 is partially confirmed. Also the odds to churn for households who earn a higher income are 1.185 times higher compared to households with a medium income level.

Several household specific independent variables, namely household size and social class, appear to be insignificantly differencing on all levels. This means that the data is neither able to confirm nor reject hypotheses five and six.

32 healthcare client, and Compliant healthcare client (1.264, 1.239, and 1.288 times larger respectively compared to the base category). As table 8 shows there are multiple combinations where the odds to churn significantly differ. Values (odd ratios) < 1 indicate that the category is less likely to churn by a factor as indicated in the corresponding cell. Visa versa, values (odd ratios > 1) indicate that the category is more likely to churn by a factor as indicated in the corresponding cell. Table 8 shows that three mentality groups (Idiosyncratic-, Quality oriented-, and Pragmatic healthcare client) are significantly more likely to churn compared to three other reference groups; Luxurious oriented-, Social critical-, and Compliant healthcare client. This indicates the importance of the underlying factors of mentality group specification with regards to churn probabilities.

Table 8 Significant odd changes to churn between mentality groups

Significant differences between mentality groups on probabilities to churn

Reference category: Signi

fi cant ly d if fer en t w .r .t .: (1 ) Co n su m p tio n o rie n ted h ea lt h ca re cli en t (2 ) Id io sy n cra ti c h ea lt h ca re c li en t (3 ) Co n v en ien ce o rien ted h ea lt h ca re cli en t (4 ) Qu ali ty o rien ted h ea lt h ca re c li en t (5 ) Lu x u rio u s o rie n ted h ea lt h ca re c li en t (6 ) S o cial crit ica l h ea lt h ca re c li en t (7 ) P ra g m ati c h ea lt h ca re c li en t (8 ) Co m p li an t h ea lt h ca re c li en t

(1) Consumption oriented healthcare client -

(2) Idiosyncratic healthcare client -

(3) Convenience oriented healthcare client -

(4) Quality oriented healthcare client - 0.797 0.815

(5) Luxurious oriented healthcare client 1.293 1.268 - 1.264

(6) Social critical healthcare client 1.267 1.243 - 1.239 (7) Pragmatic healthcare client 0.800 0.817 -

(8) Compliant healthcare client 1.317 1.292 1.288 -

Numbers are changes in odd ratios (exp(B))

Following the positive sign (exp(B) > 1) shows that the odds to churn for pragmatic healthcare clients are 1.320 times as large as than the odds to churn for compliant healthcare consumers. A one-way ANOVA was conducted to compare the effects of the different healthcare specific mentality groups on churn behavior. The ANOVA is significant which means that the different mentality groups differ significantly on churn between groups [(F(1, 11174) = 12.013, p = .001)], which supports hypothesis 7.

5.2.2 Purchase behavior