Churning behavior in the liberalizing Dutch health care market.
Master Thesis MSc Marketing Management / Intelligence
By Mendel Koornstra Student number: S3536300 XXXXXXXXXX XXXXXXXXXX XXXXXXXXXX XXXXXXXXXX
Supervisor: Prof. Dr. J.E. Wieringa Second assessor: Dr. A.E. Vomberg University of Groningen
Abstract
The liberalization of a service market has some profound consequences for both the service providers and customers active in the market. After liberalization service providers face new competitors and must address customer churning. In this research (drivers of) churning behavior in the liberalized Dutch health care market are studied. A logistic
regression followed by a latent class regression are used to create insights about factors that drive churning behavior of the customers in the liberalizing market. First the logistic
regression tested the influence of push, pull, mooring and sociodemographic factors on churning behavior of the customers. Followed by the latent class analysis that tested if these effects differ for underlying latent segments.
The results of this research showed that the push, pull, mooring and
sociodemographic factors indeed influence the churning behavior of customers in the liberalizing Dutch health insurance market. But three segments were found for which these relationships differ. A large PPM factors sensitive segment (56%) consisting of customers whose churning decisions are strongly influenced by the different push, pull, mooring and sociodemographic factors. A freedom takers segment (21%) which churning decisions are difficult to influence and a commitment sensitive segment (23%) which customers are less likely to churn when the length of the relationships grows over time.
Keywords: logistic regression; latent class regression; push, pull and mooring factors; churning behavior; switching behavior; Dutch health insurance market; liberalizing market
Acknowledgement
With writing my final thesis, my time as a student comes to an end. I started with the bachelor Commercial Economics in Enschede and it end circa eight years later as a student MSc Marketing Management & Intelligence in Groningen.
Therefore I would like to thank people who supported me during my time as a student and while writing my master thesis. First of all, I would like to thank Prof. Dr. Jaap Wieringa for his supervision during my thesis. I would also like to thank my friends and family who supported me, not only during my master thesis but also during my entire period as a student. Lastly, I want to thank my fellow students Herre Zonderland, Guus van der Veen and Tjalling de Jong for the fun time at the University of Groningen.
Table of content
Acknowledgement 2 Table of content 3 1. Introduction 5 2. Theoretical framework 8 2.1 Churning behavior 8 2.2 Push factors 9 2.2.1 Satisfaction 9 2.2.2 Quality 10 2.2.3 Trust 10 2.2.4 Commitment 11 2.2.5 Price perceptions 11 2.3 Pull factors 11 2.3.1 Alternative Attractiveness 12 2.4 Mooring factors 122.4.1 Attitude towards switching 12
2.4.2 Subjective norms 13
2.4.3 Switching costs 13
2.4.4 Prior switching behavior 14
2.4.5 Variety seeking 14
2.5 Sociodemographic factors 15
2.6 Churning behavior in liberalizing markets 16
2.7 Conceptual model 18 3. Methodology 19 3.1 Dataset 19 3.1.1 Dependent variable 19 3.1.2 Explanatory variables 19 3.2 Descriptive statistics 20 3.3 Data preparation 21 3.3.1 Outliers 22 3.3.2 Missing values 22 3.4 Research methods 23 3.4.1 Logistic regression 23
3.4.2 Latent class analysis 24
4.1.3 Validation 28
4.2 Latent class regression analysis 30
4.2.1 Number of segments 30
4.2.2 Significant relationships per segment 31
4.5 Hypotheses 34
5. Discussion 35
5.1 Effect of push, pull, mooring and sociodemographic factors 35
5.2 Effect for different segments 37
5.2.1 Commitment sensitive segment 37
5.2.2 PPM factors sensitive segment 37
5.2.3 Freedom takers segment 38
6. Conclusion 39
6.1 Findings 39
6.2 Managerial implications 40
6.3 Limitations and future research 41
References 42
Appendix 1, Hypotheses 46
Appendix 2, Logistic regression 47
Appendix 3, Latent class regression 49
Appendix 4, Descriptive statistics 52
1. Introduction
At the beginning of 2006, the Dutch health insurance industry was reformed. Dutch health insurance companies were liberalized, with the aim of allowing health insurers to compete with each other (Rosenau and Lako 2008). Facing new competition means that companies have to deal with customers that churn to other health insurance companies. Churning customers could have a major influence on companies. De la Llave, López and Angulo (2019) appoint negative consequences of churning customers. First of all, it
influences their sales revenue. Replacing churned customers entails high costs. Earlier studies showed that attracting new customers cost around five to six times more than retaining
current customers (Soeini and Rodpysh 2012). Because attracting new customers entails higher costs than investing in retaining current customers, it is financially attractive for companies to focus on retaining customers. Second, the churning of customers has a negative effect on the reputation of companies and the brand image (Saradhi and Palshikar 2011). The churned customers may influence potential customers by giving negative feedback about the company (Verbeke et al. 2012). Potential customers are then influenced by churned
customers. Word of mouth and churning decisions of customers possibly create a chain process which makes it more difficult to replace churned customers by new customers (Pinheiro and Helfert 2010). Eventually a decrease of the retention rate of customers may affect the business of a company in a negative way (Van den Poel and Lariviere 2011; Tsai and Lu 2009).
Retention programs can be used to prevent customers from churning and increase retention rates. After identifying customers with the greatest churn opportunities, they can be selected for a retention program (Holtrop, Wieringa, Gijsenberg and Verhoef 2017). As part of this retention program, marketing can be used to persuade the group of customers to stay (Morik and Köpcke 2004). The goal of these programs is to prevent customers from churning and retain customers for the company.
To gain a better understanding about why customers churn to other service providers, lots of research was conducted in the last few years. These studies can be used as input for retention programs. Bansal, Taylor and James (2005) provided a framework for factors that
survey data to test the applicability of the push, pull and mooring (PPM) migration model for predicting churning behavior. They found that the push, pull and mooring model performed better than alternative models in predicting churn behavior of customers.
Bansal, Taylor and James (2005) tested the applicability of the framework with survey data obtained within the auto repair and hair styling sector. These are markets where competition is allowed for a long time. However, in liberalized markets other factors play a strong role in influencing churn behavior than in markets that were not monopolistically from origin. Former monopolists have the competitive advantage of being well-known companies with very loyal (forced) customers who are not used to having the freedom to churn to competitors (Wieringa and Verhoef 2007). Wieringa and Verhoef (2007) found in their research, about churn behavior in liberalizing markets, 4 segments for which effects on churn behavior differed. These different effects were derived by the fact that these markets were monopolistic from origin.
Therefore, to fully understand the effect of push, pull and mooring factors on
churning behavior within a liberalized health insurance branch, there has to be accounted for a possible heterogeneous effect. This research will contribute by bridging the gap between the push, pull and mooring model and the applicability within liberalized markets. When the relationship between push, pull and mooring factors and churn behavior is not well understood within liberalizing markets, this may have a negative impact on retention
programs. Not selecting the optimal group of customers to target may have a negative impact on the effectiveness of retention campaigns. A possible consequence of less effectiveness of these campaigns are shorter relationships with customers. Relationship building is related to the success or failure of a company (Lejeune 2001). Because of that, not understanding drivers of churn behaviour in the right way may have a detrimental impact on the success or failure of a company. Based on this research problem, the following question is formulated:
Research Question: Is there a heterogeneous effect of push, pull and mooring factors
on the churning behavior of customers in a liberalizing service market?
To answer this research question, a dataset with data of a health insurance company is used. The dataset contains variables that provide insights into the churn behavior of current and churned customers. To research the influence of the push, pull and mooring factors on the churning behavior of customers, a logistic regression is applied. In addition to the logistic
regression, a latent class analysis is performed. This analysis gives us interesting insights for different effects for different latent segments.
In this research the terms churning and switching will be used as synonyms. Furthermore, the research is structured as follows: the second chapter elaborates on the theoretical framework. In this chapter the push, pull and mooring factors are discussed. Also there is explained why there may be different effects for underlying segments. Next, the methodology of the research is described. It gives insights into the data and the research methods that are used. This is followed by the results of this research. Subsequently in chapter five this research is discussed. The last chapter finalizes with conclusions, recommendations, managerial implications and guidelines for future research.
2. Theoretical framework
Chapter two lays the theoretical basis of this research. The dependent and explanatory variables are explained and substantiated with relevant literature. The effects of push, pull and mooring factors on churning behavior of customers are discussed. There also is
elaborated on why there is a possible different effect of the push, pull and mooring factors on customer churning behavior for different segments.
2.1 Churning behavior
First, the different types and importance of churning behavior is discussed. Lejeune (2001) states that customer-oriented management and building relationships with customers are factors to which success and failure is closely linked. Customer-oriented management and building relationships with customers are important aspects of customer relationship
management. Customer relationship management therefore plays an important role in the survival of a company. An important aspect of customer relationship management is the collection, preservation and use of data. Which provides insights in customer acquisition, retention, extension and selection (Lejeune 2001). As stated in chapter 1, this research will focus on the subject of customer retention.
Close related concepts to this subject are switching, lapse, churn and customer
loyalty. Switching, lapse and churn are concepts that define the percentage of customers that stop being a customer at a company (Lima, Mues and Baesens 2009; Eling and Kochanski 2013). Customer loyalty and retention is the opposite of customer churn (Wieringa and Verhoef 2007; Lejeune 2001). Central to all these concepts is maintaining or ending short- or long-term relationships between customers and companies.
There are three different types of churning. The first one is called active or deliberate churn. The reasons for this type of churn are dissatisfaction with the quality, high costs, bad support, etc. Because if customers are not satisfied, they switch to another provider or stop buying this type of product. The second type of churn is called rotational or incidental churn. Reasons for this type of churn are circumstances that prevent a customer from further
requiring the service. Examples are financial problems or a chance in the geographical location of the customer. The third type of churning is called passive or non-voluntary churn.
This occurs for instance when a customer dies or when the company ends the relationship (Lazarov and Capota 2007).
The choice of ending the relationship with non-voluntary churners is from the company itself, this group is not seen as interesting. For rotational churners, the cause of churning often lies with factors where a company has little or no influence on. For deliberate churners this is different. Their churning behavior is often driven by factors within the control of the company. For example, the price of the product or the quality of the service. That is why it is very interesting for companies to focus on this group of churners. By getting a better understanding of their churning behavior, the marketing department of a company can adapt and improve their marketing strategies and retention campaigns. What should lead to a decrease of the churn rate of customers (Lazarov and Capota 2007).
2.2 Push factors
Originally, the push, pull and mooring framework is used to explain geographical migration flows of individuals. Bansal, Taylor and James (2005) apply this model on “migration” of customers to new service providers. With this, they provide a framework for factors of customer churning behavior. Within the original model which tries to clarify geographical migration flows, push factors are defined as factors that motivate people to leave an origin (Stimson and Minnery 1998). These factors are supposed to have a negative influence on the quality indicators of life (Moon 1995). Therefore, push factors influence migration decisions of individuals. Bansal, Taylor and James (2005) argue in their paper that there is a correspondence between push factors from the migration literature and factors that influence churning behavior of customers in a service market. They argue that push factors that drive churning behavior are satisfaction, quality, trust, commitment and price
perceptions.
2.2.1 Satisfaction
Cronin, Brady and Hult (2000) describes satisfaction as an evaluation of an emotion. This evaluation of the emotion reflects the positive feelings that are evoked by the possession and/or use of a service. Customer satisfaction is often used by firms to gain customer
experiences a negative emotional evaluation about how the needs of the customer are met. In comparison with a satisfied customer, a dissatisfied customer is more likely to consider alternatives of competitors (Anderson and Mittal 2000). Which eventually may lead to churning. Based on this line of reasoning the following hypothesis is defined:
H1: The probability that a customer will churn to another service provider is higher when the
satisfaction with the current service provider is lower.
2.2.2 Quality
The second push factor addressed is service quality. Quality of a service can be defined as the extent to which a service meets or exceeds customer expectations (Zeithaml and Parasuraman 1996). Setó-Pamies defines it in a more recent research (2012) as the customers assessment of the overall excellence or superiority of the service. Shin, Chui and Lee (Shin, Chiu and Lee 2019) argue in their paper that when perceived quality is higher, the overall satisfaction is higher. Which in their turn affect customer loyalty and churning behavior in a positive way.
H2: The probability that a customer will churn to another service provider is higher when the
perceived service quality is lower.
2.2.3 Trust
Moorman, Deshpande and Zaltman (1993) define trust as a willingness to rely on an exchange partner in whom one has confidence. Various studies argue about a relationship between trust and churning behavior of customers. Moorman, Deshpande and Zaltman (1993) reason that trust is a relevant factor in situations when uncertainty plays a role. Customers may feel vulnerable in situations like economic downturn. In these situations, trust can help to reduce uncertainty and feeling safe by the decisions made by their current service provider. Thus, a lack of trust may be an important indicator of customer churning behavior in difficult times. This leads us to the following hypothesis:
H3: The probability that a customer will churn to another service provider is higher when the
2.2.4 Commitment
Commitment can be defined as a desire to maintain a valued relationship (Moorman, Zaltman and Deshpande 1992). Commitment consists of three components. First, an
instrumental component as a form of investment. For example, the costs of being insured. Second, an attitudinal component that contains an affective commitment or psychological attachment. Third, a temporal dimension indicating that the relationship exists over time (Gundlach, Achrol and Mentzer 1995). When a customer has a desire to maintain attached to these components, it will affect the length of the relationship. Therefore, when a customer is committed to a service provider this may influence the loyalty of the customer in a positive way. Based on this, the following hypothesis is stated:
H4: The probability that a customer will churn to another service provider is higher when the
commitment to the service provider is lower.
2.2.5 Price perceptions
Devaraj, Matta and Conlon (2001) found a relationship between price and customer loyalty. This is caused by the bias that consumers perceive a higher price as an extrinsic signal of quality (Kaura, Durga Prasad and Sharma 2015). Which will eventually lead to a higher customer loyalty and churn behavior. Another explanation that Kaura, Durga Prasad and Sharma (2015) give is that when a customer perceives the price as fair, they develop positive feelings against the service provider. Which will eventually affect the behavioral intentions. Based on this argumentation the following hypothesis is stated:
H5: The probability that a customer will churn to another service provider is higher when the
perceived price is higher.
2.3 Pull factors
Within the original model which tries to clarify geographical migration flows, pull factors are defined as factors that attract individuals to a geographical location (Moon 1995). Dorig and Tobler (1983) confirm this definition and describe these factors as attributes of
2.3.1 Alternative Attractiveness
Bansal, Taylor and James (2005) purpose in their research one possible pull factor: alternative attractiveness. Alternative attractiveness include customer perceptions regarding the extent to which viable competing alternatives are available in the market (Jones et al. 2000). Jones et al. (2000) provides arguments why alternative attractiveness influences churn behavior. They state that when there is less alternative attractiveness, the perceived benefits of churning by the current customer should be relatively low. Which results in a higher level of retention at their current provider. This also works the other way around. When there are more alternatives available, dissatisfied customers perceive more benefits of churning. Therefore, the following hypothesis is stated:
H6: The probability that a customer will churn to another service provider is higher when
the alternative attractiveness of competitors is higher.
2.4 Mooring factors
It often occurs that when an individual is in a situation where push and pull factors are strong, individuals do not migrate. Lee (1966) argues that this is due to contextual or
situational factors. According to Bansal, Taylor and James (2005) a model that contains only push and pull factors is too simple and does not capture the complexity of migration
decisions. To make a more complete model, mooring effects are added. Mooring factors are intervening obstacles or opportunities, like family at their current geographical location or costs that come with moving, which possibly affect the individual’s migration decision (Boyle et al. 1998). In a customer churning context for service providers, possible mooring effects are switching costs, variety seeking behavior, attitudes toward switching, subjective norms toward switching and the consumers past switching behavior.
2.4.1 Attitude towards switching
The same as for a migration decision it is argued that the attitude towards switching influences the actual behavior. In earlier research it is suggested that a favourable attitude will increase the motivation to perform a behavior when an individual perceives a high degree of controllability on their part in performing the behavior (Bansal and Taylor 2002). Therefore, in a situation where a customer had total control over the choice to switch or stay, a positive attitude toward switching would result in a strong intention to switch. Jones,
Mothersbaugh and Beatty (2000) argue a similar relationship. They argue that a higher level of dissatisfaction increases the probability of switching if switching barriers are low, instead of high. Based on this line of reasoning the following hypothesis is stated:
H7: When the attitudes of consumers toward switching are less favourable the probability
that a customer will churn to another service provider is higher.
2.4.2 Subjective norms
In marketing literature, it is often suggested that an attitude only turns into a behavioral intention when it is supported by the social environment. Grube and Morgan (1990) tested a model of attitudes and perceived social support. They found that attitude is more likely to turn into behavior when it is supported by the social environment like friends and family. Bansal and Taylor (2002) argue that this plays an important role in a service setting. Customers seem to rely more on subjective norms because of the credibility element of the service. Customers must have confidence that the service provider will deliver as expected or agreed in advance. Especially in, for example, markets for auto repair or medical services it is challenging to evaluate the service. This is because the services provided are difficult to understand for most customers. Therefore, the opinion of others may play a role making the decision to churn or stay loyal. Based on this possible relationship the following hypothesis is stated:
H8: When subjective norms toward switching are less favourable for churning behavior, the
probability that a customer will churn to another service provider is lower.
2.4.3 Switching costs
Based on the gap between intended behavior and actual behavior there may be issues out of customers control that prevent the customer from churning. One of the possible explanations for this difference are switching costs (Jones, Mothersbaugh and Beatty 2000). Switching costs include factors such as transaction costs, search costs, learning cost,
emotional costs and so on (Bansal and Taylor 2002). Bansal and Taylor (2000) argue that when these costs increase the intention to switch may be inhibited in performing the
new airplane and may experience a high level of stress when searching for the new boarding gate. Because the investment is not worth the profit the customer does not actually switch. But if these switching costs were low, the customer may actually switch flights. Therefore there can be assumed that switching costs may affect the actual churning behavior. The following hypothesis is stated:
H9: When switching costs are higher the probability that a customer will churn to another
service provider is lower.
2.4.4 Prior switching behavior
Ganesh, Arnold, and Reynolds (2000) argue in their research that prior switching behavior influences the loyalty of customers. The lack of experience with services of competitors may contribute to the way switching costs are perceived. Thus, customers that did not churn before lack experience with competitor services and therefore possibly are more likely to perceive the switching costs as higher than customers that churned before. As a consequence, customers that did not churn before have a higher probability to stay loyal as a customer, even when the customers are dissatisfied. This is in line with Bansal, Taylor and James (2005) who state that the intention of switching is related to a customer's past churning decisions. Which could be clarified by a difference in propensity for customers to seek variety in service experience. Therefore, the following hypothesis is stated:
H10: When consumers have not switched in the past the probability that a customer will
churn to another service provider is lower.
2.4.5 Variety seeking
The extent to which a customer is looking for variety is also a factor that may influence churning behavior. In some cases, customers want to experience different types of the products. Wants and needs are in that case not fulfilled by one product. Therefore, a portfolio of products is needed. This could lead to a history involving consistent churning between products. The needs of a customer are then satisfied if they balance between a right collection of products. One product fills one component of their need and another product another component (Lattin and McAlister 1985). Therefore, there can be concluded that a variety seeking customer is likely to churn between products to fill different components of their need.
Based on this argumentation the following hypothesis is stated:
H11: When consumers are less variety seeking the probability that a customer will churn to
another service provider is lower.
2.5 Sociodemographic factors
Prior research also indicates a relationship between different socio demographics and churn behavior of customers. A relation between the age of the customer and their churning behavior is possibly caused by the change of economic status (De la Llave, López and Angulo 2019). Over time, as customers reach a higher age and have a better economic status, the financial note to change service provider may decrease. Therefore, the following
hypothesis is stated:
H12: When the age of the customer increases the probability that a customer will churn to
another service provider decreases.
Beside the age of the customer, their level of education may also play an important role in their churning behavior. Education could deeply affect the individual work opportunities (Manna, Ciasullo, Cosimato and Palumbo 2017). As a result, it influences the income level of individuals because better work opportunities will eventually result in a higher income. Again, the financial note to change service provider is lower with a better economic status. Based on this line of reasoning the following hypothesis is stated:
H13: When the education level of the customer is higher the probability that a customer will
churn to another service provider is lower.
There is also strong evidence that gender influences the churning behavior of customers. A higher loyalty for women than men is based on the fact that women tend to strive more to establish and maintain relationships with people in social context and that they may do the same for relationships with companies (Melnyk, Van Osselaer and Bijmolt 2009).
Therefore, the following hypothesis is stated:
H14: When a customer is a woman the probability that a customer will churn to another
service provider is lower.
2.6 Churning behavior in liberalizing markets
Liberalizing markets have special characteristics which may influence churn behavior. Before liberalization, companies in the market were monopolists. Because these government-induced monopolists were active in markets related to basic necessity goods, not purchasing these products is not an option. In this situation churn only occurred in personal situations of customers such as death, debt or moving to another geographical area. Wieringa and Verhoef (2007) provide in their research four arguments to explain why liberalization of markets could lead to different effects for segments on churning behavior.
First, customers only have experience with the service provided by the monopolist. As argued earlier in this research, the lack of experience with services provided by competitors may contribute to the way switching costs are perceived. Customers that did not churn before lack experience with competitor services and therefore possibly have a higher probability to perceive the switching costs as higher than customers that churned before. Therefore, less customer knowledge about competing suppliers has a direct negative effect on churning (Ganesh, Arnold and Reynolds 2000).
Second, inertia possibly represents an important determinant of customer churning. Customers in liberalizing markets are originally not used to the possibility of switching between service providers. Wieringa and Verhoef (2007) argue that the inexperience of churning may lead to habitual behavior of not churning.
Third, having a long-term relationship strengthens the commitment and trust between the customer and service provider (Dwyer, Shurr and Oh 1987). It is assumed that an increase of commitment and trust is related to a lower probability of churning (Bansal, Taylor and James 2005). Monopolistic firms are not known for their excellent product or service quality or their customer orientation (Wieringa and Verhoef 2007). Because customers of former monopolists are forced loyal customers, commitment and trust to the supplier might be less influential. So churning decisions may be less driven by relational variables as commitment and trust but more by factors as inertia.
Fourth, customers may find it difficult to assess a fair price. Due to the lack of experience in the market. Because of that price may be less important. But new entrants might enter markets with an aggressive price policy, which may increase the influence of price (Wieringa and Verhoef 2007).
Wieringa and Verhoef (2007) found that the effect for the relationship between quality, perceived switching costs and attractiveness of switching on customer churning differs between 4 different segments. They distinguish segments which consist of
relationship-inertia customers, relationship-oriented customers, alternative seekers highly aware of switching costs and disloyals. Based on this, there can be concluded that treating customers in liberalizing markets as homogenous when studying the effect of push, pull and mooring factors on customer churning could lead to a misrepresentation of the true
relationship. To fully understand customer churning behavior there must be accounted for the presence of underlying latent customer segments. Therefore, the following hypothesis is tested during this research:
H15: The effect of the different push, pull, mooring and sociodemographic factors on the
probability that a customer will churn to another service provider is heterogeneous for liberalized markets.
2.7 Conceptual model
In order to create a complete picture of the hypothesised relations between push, pull, mooring and sociodemographic factors and churning behavior, they are presented in a
graphical conceptual model, see figure 1. The model also captures the possible heterogeneous effects for the hypothesised relationships.
3. Methodology
Chapter 1 and 2 laid the theoretical foundation for this research. In chapter 3 the research approach is described. There is elaborated on the possibilities offered by the used dataset, descriptive statistics to get an idea of the population the dataset contains, the required data preparations and the research methods used for model building.
3.1 Dataset
The dataset used for this research has to meet certain requirements. First, it should be able to test the effect between push, pull and mooring factors and churning behaviour of customers. It must also be possible to research different effects that are possibly caused by the liberalization of the market. Therefore, a dataset of a Dutch health insurance company is used for this research. The Dutch health insurance industry was liberalized in 2006 (Rosenau and Lako 2008), which makes the data perfectly fit the research subject.
The dataset contains 1.305.974 observations (before data cleaning) of current and churned customers. Furthermore, the dataset consists of 31 variables (before data cleaning). These variables provide insight about the characteristics, circumstances and behavior of the current and churned customers.
3.1.1 Dependent variable
The dependent variable in this research is indicated as ‘the probability that a customer will churn’ and will not stay loyal as a customer. The moment that is indicated as the moment of churn, is when a customer ends the relationship with their current insurer and leaves the company as a customer. This is in line with Lima, Mues and Baesens (2009), who describe churn as a regular customer who abandons a relationship.
3.1.2 Explanatory variables
Unfortunately, not all necessary push, pull and mooring variables are available for testing the hypotheses. Due to these data restrictions, only data is available for the following variables: satisfaction, commitment and prior switching behavior. For these variables proxy
For the push factors discussed in chapter 2 of this research three proxies are used during model building. To test the effect of commitment on the probability that a customer will churn two proxies are used. This concerns the variables ‘relatie.duur’ (Relationship length) and ‘Pakket.AV.2011’ (Additional insurances). According to Moorman, Zaltman and Deshpande (1992) commitment is a desire to maintain a valued relationship. Therefore, a longer relationship of a customer can be used as an indication of the level of commitment. Beatty, Homer and Kahle (1988) argue that when a customer is more involved with a company, he or she is more committed to the brand. Based on this, having additional insurances (higher involvement with the company) is used as an indication of commitment. Because a measure of satisfaction is not available in the dataset, klachten.totaal (total number of complaints) is used to test the relationship between satisfaction and churning behavior. Karatepe (2006) suggests a link between customer satisfaction and the total number of complaints.
In chapter 2 alternative attractiveness is suggested as the only pull factor that influences churning behavior. Unfortunately, because there is no data available on competitors or customer opinions and perceptions, it is not possible to test this possible relationship.
For the different mooring factors proposed in chapter 2 of this research one proxy is used. The effect of prior switching behavior on the probability of churning is tested with the variable overgestapt.voorheen (churned before). Customers that did not churn before lack experience with competitor services and therefore have a higher probability to perceive switching costs as higher than customers that churned before. Switching costs on their turn affect churning behavior (Ganesh, Arnold and Reynolds 2000). Therefore, whether a customer has switched before is used to measure the effect of prior switching behavior on churning behavior.
In addition to the push, pull and mooring factors also the effect of sociodemographic factors on churning behavior is tested. In chapter 2 is reasoned why age, gender and
education possible effect churning behavior of customers. These relations are tested with the variables ‘Leeftijd’, ‘Geslacht’ and ‘Opleiding’.
3.2 Descriptive statistics
In order to gain insights about the data and population within the dataset, this paragraph briefly discusses relevant characteristics of the individuals the dataset contains.
Variables that receive extra attention are sociodemographic factors, total number of complaints, relationship length and prior switching behavior. Within the described data, missing values and outliers are dealt with and replaced or removed if necessary.
After cleaning and preparing the dataset, it contains a total of 927.992 observations and 10 variables. Every observation is linked to a ‘persoonsnummer’ (personal number). Every observation is a unique current or churned customer. Of all observations, 880.401 are still customers and 47.591 customers have churned to another health insurer. For each observation it is known what the duration of the relationship is or was and whether a customer is churned. Therefore, it is known how long a relationship lasted per churned customer.
The average age of a customer in the dataset is 42 years. Of all current and churned customers, 447.725 are men and 480.267 are women. Looking at the variable ‘opleiding’ (education), customers are divided into 3 categories. Category 1 contains customers with a lower education, category 3 customers with a high education. Of all customers 392.049 fall in category 1, 296.326 in category 2 and 239.617 in category 3.
On average, customers were 1.6 times in contact with the insurer. Of all customers 426.670 never had contact with the health insurer. 485.414 customers have contacted the helpdesk between 1 and 10 times. This means that 45% of the customers have never had contact and 53% contacted the insurer between 1 and 10 times. Of these contacts, 3.369 were customers with a complaint. In most cases customers had 1 complaint, 3.070 times. 255 customers had 2 complaints. In 44 cases a customer had 3 complaints. The average number of complaints per customer is therefore 0.004.
The average relationship length of a customer is 16 years. Approximately 75% of the (former) customers have a relationship of 22 years or less. 35.109 of the customers have a relationship less than one year. The longest relationship a customer has within the dataset is 94 years.
Of all customers, 20% did not use any additional insurance packages. 80% made use of 1 of the 17 different types of additional insurance packages.
The churn behavior in the dataset is divided into 9,8% that did not churn before they became customers, the other 90,2% appear to have churned before. Of all the 1.226.167
3.3 Data preparation
To prepare the data for analysis, there are required steps that must be followed. For example, outliers and missing values must be handled. Data points that are labelled as outlier differ quite heavily from the rest of the observations (Wickham and Grolemund 2017). These data points are so far from the cluster of the other data points that they can be harmful to the research. Not treating missing values in the right way may lead to biased estimates, distorted statistical power, and invalid conclusions (Acock 2005). This chapter explain how this research deals with missing values and outliers. Before outliers and missing values are treated, the original obtained dataset is cleaned up by removing all irrelevant variables from the dataset.
3.3.1 Outliers
Like missing values, outliers can be a threat to the reliability of the research. According to Breunig, Kriegel and Sander (1999) outliers are objects deviating from the major distribution of the dataset. An outlier can be identified by localizing observations that fall 1.5 times outside the interquartile range. This applies to both above the upper quartile and below the lower quartile (Leeflang et al. 2015). However, there is no data point within the dataset that falls outside the lower or upper quartile, this can be assumed as an unrealistic value that may harm the reliability of the research.
3.3.2 Missing values
The original dataset contains several missing values. These data fields that provide no useful data are filled with the values "9999". To not violate the reliability of the research, these values are replaced or removed.
For the variable ‘relatie.duur‘ (relationship length) there are 58,723 observations with the value "9999". These values can be assumed to be impossible for the length of the
relationship. Therefore, they are classified as missing value. These observations are removed from the dataset.
The variable ‘Opleiding’ (education) contains 21.084 observations with the value “9999”. Because the variable ‘Opleiding’ is a categorical variable which categorizes observations in categories from 1 to 3, this value is not valid. Therefore, these values are deleted from the data set.
The variable ‘Pakket.AV.2011’ (additional insurances) contains 298.175 observations with the value “9999”. The variable ‘Pakket.AV.2011’ is a categorical variable. Additional insurances are categorized with values from 0 to 17. A value “9999” is therefore classified as a missing value and deleted from the dataset.
3.4 Research methods
In the first part of model building, a logistic regression is used. It is used to test the relationships between the push, pull, mooring and sociodemographic factors and the churning behavior of customers. In part two of the model building process a latent class regression analysis is used. During the latent class regression analysis, it is tested how different underlying segments influence the relationship between the push, pull, mooring and sociodemographic factors and the churning behavior of customers.
3.4.1 Logistic regression
To predict the probability of churning for current customers, a binary logistic
regression is used. A binary logistic regression is suitable to use because it takes into account that the dependent variable is categorical. The influence of several independent variables on the dependent variables can also be included in the model (Tranmer and Elliot 2008). The relation between choice and the explanatory observable indicators is represented in the following equation.
Where εi is the value of the random disturbance term. Explanatory observable variables are
represented by xi. In this equation y* is a latent variable. This y* can be interpreted as the
difference between unobserved preference between stay loyal as a customer or churn to another health insurer. Eventually the Y has to be interpreted, which is the probability that a customer will churn or not. The relationship between Y and y* can be specified as followed.
For using a binary logistic regression there are a few assumptions that need to be checked. First of all, it is assumed that the dependent variable is binary and ordinal, which assumption is met in the case of churn or no churn dependent variable.
The second assumption that must be checked is about the sample size (Starkweather and Moske 2011). Leeflang et al. (2015) indicated that between the 5 and 10 observations are required per variable. The dataset that is used during this study contains 927.992
observations. Which demonstrates that this assumption is met.
The third assumption that must be checked is the absence of multicollinearity.
Multicollinearity is a statistical phenomenon that occurs when explanatory variables used for a logistic regression model are strongly correlated. Multicollinearity can mainly be detected using the variance inflation factor, also known as VIF score (Midi, Sarkar and Rana 2010). This assumption will be further discussed in Chapter 4.
The measures used to validate different models are the hit rate, top decile lift, Gini coefficient and the mean squared error. These measures are further discussed in chapter 4.
3.4.2 Latent class analysis
Based on theory there is assumed that there may be underlying latent classes that influence the effect between the dependent and independent variable. Therefore, a Latent Class Analysis (LCA) is performed. The basic idea of the latent class analysis is that some of the variables of a statistical model differ across unobserved subgroups (Vermunt and
Magidson 2004). Based on a latent class regression, the posterior probability is used to segment the observations in the dataset. The observations are assigned to the class with the maximum posterior probability (Leisch 2004). The posterior probability that an observation belongs to a class is given by the following equation:
After assigning observations to a class, parameters are estimated. The log-likelihood of a sample of observations is given by:
It usually cannot directly be maximized. For this the iterative EM algorithm is used. First the posterior class probability for each observation is estimated and the prior class probability is derived (Leisch 2004). After that, the log-likelihood for each component is maximized using the posterior probabilities as weights:
The estimation and maximizing steps are repeated until the likelihood improvement falls under a pre-specified threshold or a maximum number of iterations is reached (Leisch 2004).
The ideal number of segments is decided based on the information criteria AIC and BIC. This is further discussed in Chapter 4.
4. Results
In chapter 4 the results of this research are discussed. First, there is elaborated on the results of the logistic regression. Which contains the significant effects and assumptions of the model. Next, there is elaborated on the latent class regression results. The ideal number of segments is decided and the significant effects are discussed. This chapter will end with the acceptance or rejection of the hypotheses state in chapter 2.
4.1 Logistic regression
As mentioned above, the significant effects and assumptions for the logistic
regression are discussed. Discussing the significant effects consists of interpreting the p-value and calculating the odds ratio. The odds ratio will provide insights into the decrease or
increase in the probability of churning with an increase of 1 by the explanatory variable. Also, assumptions are discussed in this paragraph. In subparagraph 2.1 of chapter 3 there is already elaborated on the assumptions that have to be met. The last assumption that still needs to be addressed is checking for the absence of multicollinearity.
4.1.1 Significant relationships
Table 1 presents which relations with churn behavior are significant based on the logistic regression. Also the estimates and standard error are provided.
Table 1, parameter estimates and significant effects
Variable Estimates Std. Error P-value
(intercept) -2.3632264 0.0198734 0.005 ***
Relationship length -0.0372101 0.0007102 0.005 ***
Number of complaints 0.8022826 0.0537805 0.005 ***
Not churned before -0.3561140 0.0206653 0.005 ***
Additional insurance 0.0157264 0.0011940 0.005 ***
Age -0.0107633 0.0003330 0.005 ***
Gender 0.0460765 0.0116295 0.005 ***
Education 0.1104953 0.0072974 0.005 ***
The relationship length between the customer and the insurer has a significant (p < .005) and negative effect (exp(B) < 1) on the probability that a customer will churn. Looking at the odds ratio it is inferred that when the relationship increases within 1 year, the odds that indicates if a customer will churn decreases with 0.963.
The number of complaints a customer submits has a significant (p < .005) and positive (exp(B) > 1) relationship with the probability that a customer churns to another health insurer. By interpreting the odds ratio there can be inferred that by the increase of one complaint the odds that a customer churns increases by 2.230.
If a customer has not churned before is significant (p < .005) and negative (exp(B) < 1) related to the probability that a customer will churn. When interpreting the odds ratio it is inferred that when a customer has not churned before he or she became a customer the odds of churning decreases with 0.700.
Having additional insurance packages is significant (p < .005) and positive (exp(B) > 1) related with the probability that a customer will churn. When the amount of additional insurance packages increase by 1, the odds that a customer will churn increases with 1.016.
The age of the customer has a significant (p < .005) and negative effect (exp(B) < 1) on the probability that a customer will churn. Looking at the odds ratio it is inferred that when the age increases with 1 year, the odds that a customer will churn decreases with 0.011.
Gender has a significant (p < .005) effect on churning behavior. Being a woman has a positive effect (exp(B) > 1) on the probability that a customer churns to another health insurer. By interpreting the odds ratio there can be inferred that when being a woman instead of man increases the odds of churning by 1.047.
The level of education has a significant (p < .01) and positive (exp(B) > 1) effect on the probability that a customer will churn. When interpreting the estimate, it is inferred that when the level of education increases by 1 the odds of churning increases with 1.117.
4.1.2 Assumptions
As stated in chapter 3, the last assumption that has to be checked is multicollinearity. Multicollinearity is a linear relation between two or more variables. This problem may cause serious difficulties with the reliability of estimates of the model parameters (Alin 2010).
An overview of the VIF scores per variable are presented in table 2.
Table 2, VIF scores
Variables VIF score
Relationship length 1.406078 Amount of complaints 1.002619 Churned before 1.202449 Additional insurance 1.049291 Age 1.610035 Gender 1.013322 Education 1.005947
Based on the VIF scores in table 3, it is concluded that multicollinearity does not play a role in the estimated model.
4.1.3 Validation
An important part of model building is testing the validity of the model. Two different models are compared to test which model comes closest to reality and is best in predicting which customer will churn or stay loyal. The models that are compared are a null model and a model containing the explanatory variables discussed in chapter 3.
Measures that are used for comparing the models are the top decile lift (TDL), Gini coefficient, the hit rate and the mean squared error (MSE).
The top decile lift is a popular measure used for model validation. It gives an indication of the predictive performance for the 10% of the observations with the highest model predictions. A TDL of 1 is expected for a random model. A model with a TDL of 3 indicates that 3 times more positive cases are identified by the model than a random model would. Bose and Chen (2009) state that the higher the TDL the better the model.
Originally the Gini coefficient is used to measure income inequality (Chen, Tsaur and Rhai 1982). The Gini coefficient can also be used for measuring the inequality in data of other natures. In this research it indicates the inequality in predicted data. A Gini coefficient of 1 indicates full inequality. a lower Gini coefficient indicates no inequality.
The hit rate is about the number of right predictions a model generates when it is applied on new data. The model is built on one part of the dataset. This is called the training data. Which contains 75% of the observations. The predictive power of the model is tested on the other part of the dataset. Which contains 25% of the observations. The higher the hit rate, the better the model is in predicting the right outcome for an observation.
The mean squared error indicates the number of errors in the predictions. A low mean squared error implies a low average error in the predictions produced by the model. A higher mean squared error implies a higher average error produced. Therefore a lower mean squared error can be perceived as better.
The null model does not include any variables. The validity measures for this model are as expected, see table 3. The top decile lift for the null model is 1. Which is expected for a random model. The Gini coefficient is 0. Which indicates no inequality. For all the predicted outcomes 96% are correct (hit rate). This is because the null model does not predict any observations to churn. The 4% wrong predicted outcomes are all the observations that actually churned. MSE for the null model is 0.05138.
The PPM + sociodemographics model includes all variables as discussed in chapter 3. The top decile lift of the model is 1.83. Which is higher and therefore outperformance the null model. The Gini coefficient is 0,44. Which indicates that there is inequality in the data. The hit rate of the model is again 96%. For the same as the null model, the PPM model does not predict any customers to churn. The MSE of the PPM model is 0.04681. Which is lower and therefore better than the null model. Because on average this model produces less errors than the null model. Based on these measures it is concluded that the null model
outperformance the PPM + sociodemographics model and therefore is better in explaining the churning behavior of customers.
Table 3, validity measures
Model TDL Gini Hit rate MSE
Null model 1 0 96% 0.05138
PPM + sociodemographics model 1.83 0,44 96% 0.04681
4.2 Latent class regression analysis
First, there is decided on the ideal number of segments. For these different segments the estimates, standard error and significant levels differ. Therefore, the results per segment are discussed.
4.2.1 Number of segments
To select the ideal number of segments a broad range different latent class regression models are estimated. For each latent class regression model, results are estimated with a different number of segments. Based on the Akaike’s Information Criteria and the Bayesian Information Criteria the latent class regression model with the optimal number of segments is selected. Table 4, see next page, provides an overview of the AIC and BIC scores for
different latent class regression models. The AIC and BIC scores for models with 1 to 6 segments are included.
Table 4, AIC and BIC
Number of segments AIC BIC
1 245120 245211.6 2 245138.6 245331.9 3 243562.6 243858.3 4 243646.8 244044.8 5 245192.4 245692.7 6 245210.6 245813.2
Note: AIC = Akaike’s Information Criteria, BIC = Bayesian Information Criteria
Based on the information criteria there is concluded that 3 segments is the ideal number of segments. Both the BIC and the AIC have the lowest score for 3 segments, see table 3.
4.2.2 Significant relationships per segment
Significant effects for the three different segments are discussed in this sub paragraph. Model output is shown in table 5. The relationships between the explanatory variables and churning behavior differ between the 3 different segments. For example, it is noticeable that not all relationships, as stated in the hypotheses, are significant.
Table 5, parameter estimates and significant effects segment 1
Segment 1 Segment 2 Segment 3
Variable Estimates P-value Estimates P-value Estimates P-value
(intercept) 6.5001127 0.005 *** -2.7441180 0.005 *** -5.56069408 0.005 ***
Relationship length -0.0269790 0.005 *** -0.3506110 0.005 *** 0.00344323 0.1 Number of complaints 9.5715047 0.1 1.6718496 0.005 *** -4.42916864 - Not churned before -10.720966 0.01 * -0.7078570 0.005 *** 3.09040180 0.005 ***
Additional insurance 0.0045592 0.1 0.0688332 0.005 *** 0.00096027 1
Age -0.1119065 0.005 *** 0.0064139 0.005 *** 0.00200889 1
Gender 0.1263560 0.01 * 0.2646011 0.005 *** -0.05169841 0.1
Education 0.1297648 0.005 *** 0.0946690 0.005 *** 0.12693584 0.005 ***
Note: Signif. codes: 0.005 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’
First the significant relationships for segment 1 are discussed. The relationship length between the customer and the insurer is significant (p < .005) and has a negative effect (exp(B) < 1) on the probability that a customer will churn. Looking at the odds it is inferred that when the relationship increases within 1 year, the odds that a customer will churn decreases with 0.973.
If a customer has not churned before is significant (p < .01) and negative (exp(B) < 1) related to the probability of churning.The odds ratio inferred that when a customer has not churned before he or she became a customer the odds of churning decreases with 2.208.
The age of the customer is significant (p < .005) and negative related (exp(B) < 1) to the probability that a customer will churn. The odds ratio inferred that when the age of a customer increases with 1 year, the odds that a customer will churn decreases with 0.894.
Gender is significant (p < .01) related to churning behavior. Being a woman has a positive effect (exp(B) > 1) on the probability that a customer churns to another health insurer. By interpreting the odds ratio there can be inferred that when being a woman instead of man the odds that a customer churns increases by 1.347.
In this segment the level of education has a significant (p < .005) and positive (exp(B) > 1) effect on the probability that a customer will churn. The odds ratio inferred that when the level of education increases by 1 the odds of churning increases with 1.139.
For the relationship between the number of complaints submitted and having
additional insurance packages on the probability of churning there is not found a significant relationship.
After segment 1, the significant effects for segment 2 are discussed. The relationship length between the customer and the insurer is significant (p < .005) and negative (exp(B) < 1) with the probability that a customer will churn. The odds ratio inferred that when the relationship increases with 1 year, the odds that a customer will churn decreases with 0.997.
The number of complaints a customer submits has a significant (p < .005) and positive (exp(B) > 1) relationship with the probability that a customer churns to another health insurer. By interpreting the odds there can be inferred that by the increase of one complaint the odds that a customer churns increases by 5.322.
If a customer has not churned before is significant (p < .005) and negative (exp(B) < 1) related to the probability that a customer will churn. When interpreting the odds ratio, it is inferred that when a customer has not churned before he or she became a customer the odds of churning decreases with 0.493.
Having additional insurance packages is significant (p < .005) and positive (exp(B) > 1) related with the probability that a customer will churn. When the number of additional insurance packages increases by 1, the odds that a customer will churn increases with 1.071.
The age of the customer has a significant (p < .005) and positive effect (exp(B) > 1) on the probability that a customer will churn. Looking at the odds it is inferred that when the age increases with 1 year, the odds that a customer will churn decreases with 1.006.
Gender has a significant (p < .005) effect on churning behavior. Being a woman has a positive effect (exp(B) > 1) on the probability that a customer churns to another health
insurer. By interpreting the odds there can be inferred that when being a woman instead of man the odds that a customer churns increases by 1.303.
The level of education has a significant (p < .01) and positive (exp(B) > 1) effect on the probability that a customer will churn. When interpreting the odds ratio it is inferred that when the level of education increases by 1 the odds of churning increases with 1.099.
Lastly the results for segment 3 are reviewed. If a customer has not churned before is significant (p < .005) and positive (exp(B) > 1) related to the probability that a customer will churn. When interpreting the odds ratio, it is inferred that when a customer has not churned before he or she became a customer the odds of churning decreases with 21.985.
The level of education has a significant (p < .01) and positive (exp(B) > 1) effect on the probability that a customer will churn. When interpreting the estimate, it is inferred that when the level of education increases by 1 the odds of churning increases with 1.135.
In this segment there is not found a significant relationship between the number of complaints submitted, the relationship length, having additional insurance packages, age and gender and the odds of churning.
4.5 Hypotheses
One of the goals of this research is to find in what way push, pull, mooring and sociodemographic factors influence the churning behavior of customers in a liberalizing service market.During the research different types of push, pull and mooring factors are identified. Also is accounted for the effect of sociodemographic factors and possible latent classes for which the effect on their churning behavior differs.Table 6, see below, shows an overview of the accepted and rejected hypotheses. Hypotheses are accepted or rejected based on the logistic regression results. Accepting or rejecting hypotheses based on the latent class regression could create a distorted picture of the relationships tested. This is because the goal of the latent class regression is to find different effects for segments. Hypothesis 15 is
formulated to cover the existence of a heterogeneous effect for the tested relationships. Due to data limitations, not all hypotheses are tested. Hypotheses that are not tested are omitted from the table. Hypotheses 2, 10,12 and 15 are accepted.These relationships are found as stated in the hypothesis. For hypotheses 13 and 14 there was indeed found a relationship. But for hypotheses 13 and 14 these results were positive instead of negative. Hypothesis 4 is partly accepted because there is not found a positive relationship for all used proxies.
Table 5, hypotheses
Hypothesis Accepted/Rejected
H2 The probability that a customer will churn to another service provider is higher when the satisfaction with the current service provider is lower.
Accepted
H4 The probability that a customer will churn to another service provider is higher when the commitment to the service provider is lower.
Partly accepted
H10 When consumers have not switched in the past the probability that a customer will churn to another service provider is lower.
Accepted
H12 When the age of the customer increases the probability that a customer will churn to another service provider decreases.
Accepted
H13 When the education level of the customer is higher the probability that a customer will churn to another service provider is lower.
Rejected
H14 When a customer is a woman the probability that a customer will churn to another service provider is lower.
Rejected
H15 The effect of the different push, pull, mooring and sociodemographic factors on the probability that a customer will churn to another service provider is heterogeneous for liberalized markets.
5. Discussion
This chapter establishes the link between the theoretical framework and the results of this research. First we discuss how the push pull, mooring and sociodemographic factors affect churning behavior, based on the results of the logistic regression. Secondly we discuss how this effect differs for the three different segments, based on the results of the latent class analysis.
5.1 Effect of push, pull, mooring and sociodemographic factors
Bansal, Taylor and James (2005) argued that quality, satisfaction, trust, commitment and price perceptions are possible push factors that drive customers away from their current service provider. Therefore, they possibly could affect the customer loyalty and churn
behavior. Unfortunately, due to data restrictions only the relation between churn behavior and satisfaction and commitment is tested.
Despite these restrictions, there indeed is found a relationship between churn behavior and satisfaction and commitment. The number of complaints a customer has submitted is used as a proxy for satisfaction. There is indeed found a relationship between the number of complaints a customer submitted and their churn behavior. The more complaints a customer submitted, the higher the probability that the customer will churn to another service provider. This is consistent with the hypothesis that a lower satisfaction is negatively related with the probability of churning.
To test the relationship between churn behavior and commitment, the length of the relationship and additional insurance packages are used as a proxy. The longer the length of the relationship, the higher the commitment to the service provider. But in contrast, the more additional insurance packages the higher the probability of churning. Therefore, the
hypothesis about the relationship between commitment and churning behavior is partly accepted.
The paper of Bansal, Taylor and James (2005) provides us with one possible push factor. They argue that the attractiveness of alternative service providers could lure customers in their direction. Because there was no data available about the attractiveness of competitors,
Contextual or situational (mooring) factors that possibly affect the churning behavior of customers are switching costs, consumers propensity to seek variety, consumers attitudes and subjective norms toward switching and the consumers switching history (Bansal, Taylor and James 2005). In this research the relationship between the customers churning history and churning behavior is tested. The data provided history about the fact that a customer had churned before they became a customer or not. There is found that when a customer churned before he or she was more likely to churn again. In other words, the probability that he or she churns increases when he or she has churned in the past. This is in line with the hypothesis which suggests that the probability of churning increases when the customer has switched in the past.
In addition to the push, pull and mooring factors, various sociodemographic factors are also included in the model. The factors included are age, gender and education.
For age there is found a significant relationship. The higher the age of the customer, the lower the probability that a customer will churn to another service provider. This is in line with the stated hypothesis. In which a negative relationship between age and the probability that a customer will churn is proposed.
For gender also a significant relation with churn behavior is found. But this relation is different than expected. In chapter 2 there is argued that women are perceived to be more loyal than men, due to the fact that women tend to strive more to establish and maintain relationships (Melnyk, Van Osselaer and Bijmolt 2009). In this research it is found that being a woman increases the probability of churning. This different relationship could be caused by the fact that monopolistic firms typically are not known because of their service and product quality or their customer orientation (Wieringa and Verhoef 2007). Women are more social-relationship oriented and are likely to be more influenced by their evaluations of interactions with consulting services or for instance sales staff (Sharma Chen and Luk 2012). The original service provider (monopolist) probably is not very customer oriented. Therefore, female customers possibly, after liberalization, will look for another service provider. To find a good long-term relationship. For male customers does it not apply, because they care less for good long-term relationships.
Education also was found to have a significant relationship with churning behavior. But again this relationship is different than expected. In chapter 2 there was hypothesised that the relationship between the level of education and probability was negative. This is because education could deeply affect the individual work opportunities (Manna, Ciasullo, Cosimato
and Palumbo 2017). As a result, it influences the income level of individuals because better work opportunities will eventually result in a higher income. When a customer has a higher income the financial note to change service provider is lower. An opposite effect could be caused by the fact that educational efforts help customers understand proper usage of a service (Lee, Lee and Feick 2001). Therefore, it could be that after liberalization of the market, higher educated customers better understand the benefits of switching to another service provider than lower educated customers.
5.2 Effect for different segments
In this research there are found 3 segments for which the effects of the explanatory variables on churning behavior differ. One of the segments consist of commitment sensitive customers, another out of PPM factors sensitive customers and one with freedom takers customers. The commitment sensitive segment contains 149.920 customers. The PPM factor sensitive segment is the largest segment and contains 363.195 observations. The smallest segment is the segment with freedom taking customers, which contains 127.972 customers.
5.2.1 Commitment sensitive segment
Segment 1 consists of 23% of all customers. The segment is defined as the
commitment sensitive segment. In paragraph 6 of chapter 2 is argued why the effect of the different push, pull and mooring factors may be different in liberalizing markets. These different effects possibly are caused by the way switching costs are perceived, the role of inertia and a different influence of commitment, trust and price. The segment is called the commitment sensitive segment because, looking at the results of this research, these
customers seem not to be influenced by most push, pull and mooring factors. Inertia indeed seems to play a role in this segment.
In this segment only relationship length, which is associated with commitment, and age play a strong role in effecting churning behavior. Less strong relationships are found for gender and prior churning behavior. When a customer did not churn before, he is more likely to stay loyal. As such, the inexperience of churning may have led to habitual behavior of not
5.2.2 PPM factors sensitive segment
As stated above, the push, pull and mooring (PPM) factors sensitive segment consist of 363.195 customers. This is 56% of all customers. The possible influence of the factors, mentioned in chapter 2 paragraph 6, on churning behavior in liberalizing markets do not appear to have an influence on customers in this segment. Looking at the results for this segment, it is concluded that push, pull and mooring factors strongly influence churning behavior of the customers.
5.2.3 Freedom takers segment
This segment, which consists of 21% of all customers, is named as the freedom takers segment. This based on the fact that the only two factors that influence their churning
behavior is their prior churning behavior and their level of education. When a customer did not churn before, he or she is more likely to churn. This is found to be a strong effect. None of the other push, pull, mooring and sociodemographic factors seem to play a role in their churning decisions.
