What makes you (s)tick?

Master thesis

MSc. Marketing Management & Marketing Intelligence

What makes you (s)tick?

An empirical study on factors inducing customer switching behavior among

Dutch Health Insurance Customers.

by

Diederick Dorenbos Student Number: S3855627 2020/2021 <Adress> <Adress> d.m.dorenbos@student.rug.nl <Number>

Supervisor: Prof. Dr. J.E. Wieringa Second assessor: A. Bhattacharya

Abstract

Firms now realize that to maximize customer lifetime value, retaining customers is of paramount importance. When a firm loses a customer, it suffers from revenue loss, increased acquisition cost, and a lost opportunity to cross- or upsell.

To allocate resources to customer retention efforts, marketing managers need knowledge of two elements of customer switching behavior are key. First, the drivers of customer switching behavior must be understood. Next, accurate prediction techniques must be established.

In this research, we first address the limitations of customer satisfaction, customer engagement and customer churn intent as predictors of customer churn. Also, we examine the effect of customer characteristics. Next, we compare the predictive accuracy of conventional and machine learning modelling methods.

In the context of the Dutch Health insurance market, the following results are obtained: (1) There is no concluding evidence of the effect of economic or service-related determinants of satisfaction on likelihood to churn of dissatisfied customers, (2) older segment of people are less likely to switch than the younger segment (3) customer engagement yields a positive effect on the intent to churn by a customer, (4) logistic regression and support vector machines are the preferred techniques for out-of-sample predictions. Additional future research directions are discussed.

Key words: Customer Churn, Customer Churn Intent, Customer Satisfaction, Customer Engagement, Logistic Regression, Machine Learning Methods

Acknowledgment

This thesis is the product of the hard work I put into the past one and a half year studying marketing management and marketing intelligence at the Rijksuniversiteit Groningen. After finishing a bachelor in the field of engineering, I feel a sense of pride in successfully finishing a pre-master marketing and double-track master marketing at the Rijksuniversiteit Groningen. In retrospect, the master marketing has indeed sparked my interest in this field of research and business.

Table of Contents

1. Introduction ... 5 2. Theoretical Framework ... 8 2.1. Database Marketing ... 8 2.2. Customer lifetime ... 9 2.3. Drivers of churn... 10 2.4. Churn intent ... 14 2.5. Predictive modelling ... 15 2.6. Conceptual Model ... 16 3. Methodology ... 17 3.1. Data Description ... 17 3.1.1. Data collection ... 17 3.1.2. Data preparation ... 17 3.1.3. Variable explanation ... 18

3.1.4. Outliers and missing values... 21

3.2. Model specification ... 21

3.2.1. Binomial Logistic Regression ... 22

3.2.2. Modelling Methods ... 23

3.2.3. Mediation analysis ... 24

3.3. Model estimation ... 25

3.4. Model Comparison ... 26

4. Results ... 27

4.1. Description of the sample ... 27

4.2. Predictive modelling of churn behavior ... 28

4.2.1. Validation Churn Model ... 28

4.2.2. Logistic regression ... 29

4.3. Model performance comparison ... 32

4.3.1. Model Performance Conclusion ... 34

4.4. Mediation analysis ... 34

4.5. Churn Intent Analysis ... 35

4.6. Churn in Retrospect ... 37

5. Discussion ... 39

1. Introduction

Traditionally, the leading goal for firms has been to derive value solely from the product or services it offered to their customers. As such, decision making by marketing managers resolved around increasing sales, market share, and product margins. However, in the last decade, markets have been changing. For instance, consumers have become more value-conscious, increasingly intolerant to low-quality products and services, are exposed to customer-to-customer reviews, and more willing to switch between firms (Kumar & Reinartz, 2012). Subsequently, for firms to survive in a market that competes for increasingly conscious consumers, creating valuable relationships between customer and firm has caught increased attention by marketing managers.

As the relationship between customers and a firm has changed, firms now focus on the value of individual customers (Shah, Rust, Parasuraman, Staelin & Day, 2006). This concept differs from the traditional marketing concept in that it is focused on deriving value from the market by fulfilling the needs of individual customers (Hoekstra, Leeflang, & Wittink, 1999; Ramani & Kumar, 2008). The loyalty of customers in these relationships lead to a two-way interaction between firm and customer, in which value creation for customers also leads to value creation by customers (Shah et al., 2006). Following this principle, marketing decision making is now commonly aimed at establishing a continued and profitable relationship with the firm’s individual customers (Ascarza et al., 2017).

A common way of measuring the value an individual customer provides, is by using the customer lifetime value (CLV). The customer lifetime metric represents the total value a customer provides to the firm. It is characterized by the long-term focus of managing acquisition, retention, and development of individual customers (Blattberg, Kim & Neslin, 2008). As the CLV is often used to diagnose a firm's financial health and assist in making marketing decisions, a common goal of firms is to maximize this value. Naturally, longer-term relationships with customers are desirable as it increases the potential value each customer provides. To maximize the customer lifetime value, however, customers must be retained, and ‘customer churn’ prevented (Gupta, Lehmann, & Stuart, 2004; Venkatesan & Kumar, 2004).

even the most commonly used churn predictors in academic research, there are vast differences in the predictor's effect on customer churn likelihood. Moreover, a preferred modelling technique has yet to be established.

Firstly, the commonly used satisfaction rating has been a debated metric to use as a predictor of customer churn. Prior research suggests that customer satisfaction shares a direct relationship with the length of a customer's relationship (Bolton, 1998) and can be used to predict retention likelihood (Rust and Zahorik, 1993). However, de Haan, Verhoef, & Wiesel (2015) find significant differences in the extent to which customer satisfaction on the individual customer level can predict customer churn. These differences were found across 18 different industries in both contractual and non-contractual setting. Anderson and Mittal (2000) point out that the asymmetric link between satisfaction and retention is often disregarded. They conclude that satisfied and dissatisfied customers can have a significantly different effect on the expected retention likelihood.

Moreover, customer satisfaction in itself may not act as a unitary construct. Geyskens, Steenkamp & Kumar (1999) find that economic satisfaction and service-related satisfaction are distinct constructs with different relationships to consumer behavior and churn predictors. In our research, we focus on the limitations of the customer satisfaction metric in predicting customer churn. We recognize customer satisfaction and dissatisfaction act as distinct constructs on customer churn. Then, we use the determinants of customer satisfaction in the context of dissatisfied customers.

In addition to customer satisfaction, looking at the drivers of churn by using customer engagement variables has become increasingly relevant (Bijmolt, Leeflang, Block, Eisenbeiss, Hardie, Lemmens, & Saffert, 2010). In previous literature, the extent to which a customer is engaged with the firm and other customers often limited or left out of predictive modelling. de Haan et al. (2015) find that customer engagement is a significant predictor in 1 out of 18 researched industries on the individual customer level. De Haan et al. define customer engagement as a customer that has contacted the firm at least once.

underestimating the customer lifetime value by up to 40 percent (Wangenheim and Bayo´n. 2007).

Thirdly, previous literature has not reached a consensus on which modelling techniques perform best for predicting churn. Most commonly, logistic regression and classification tree models are used (Neslin, Gupta, Kamakura, Lu, & Mason, 2006). More advanced models as neural network, support vector machine, bagging, boosting and random forests have also not gained widespread use among marketing scientists (See Risselada, Verhoef, & Bijmolt, 2010 for a review). We compare predictive modelling performance across several conventional and machine learning methods to determine the best performing techniques.

Notably, observed customer switching behavior does not always tell the whole story. As the goal of retention management is to retain customers, we deem it particularly interesting to establish the drivers of churn intentions rather than observed churn. The observed customer switching behavior provides marketing managers with objective observations of customers' act switching to another firm. On the contrary, a customer's stated intent to switch represents the customers' self-reported likelihood of churning (Wirtz, Xiao Chiang & Malhotra, 2014). Marketing managers can use the self-reported likelihood to churn to design a proactive retention campaign targeting customers who intent to churn rather than saving those who have already churned (Ascarza et al., 2017). Hence, we distinguish between predictors of actual customer churn and customer churn intent and determine differences.

This research addresses the limitations of the commonly used predictors in customer churn. We expand the constructs of customer satisfaction, customer engagement and distinguish between customer churn and customer churn intent. We compare predictive modelling performance across several conventional and machine learning methods. Lastly, we use varying customer characteristics to test for the effect on churn likelihood. We formulated the following research question:

“What are the drivers of customer switching behavior at the customer level?”

Given our focus on the predictors of switching behavior mentioned above, this led to the following sub-questions:

Research Question 2: What is the effect of customer engagement on customer switching behavior at the customer level?

Research Question 3: What role does the intention to switch play in actual switching behavior at a customer level?

Research Question 4: Does the modelling method matter in predicting customer switching behavior at the customer level?

We contribute to marketing research in that we investigate the effect of predictors of customer churn by addressing the limitations in the literature of the satisfaction, engagement, and churn constructs. Understanding and identifying why customers stick with or leave a firm is a relevant topic of interest. It directly impacts the customer lifetime value, thus the firm's financial health (Kumar, Aksoy, Donkers, Venkatesan, Wiesel & Tillmanns, 2010). Additionally, determining the best performing models in customer churn can lead to

financially meaningful differences (Neslin et al., 2006). We aim to improve our understanding of effective churn modelling methods by comparing predictive accuracy across conventional and machine learning modelling techniques.

First, a theoretical framework is presented, followed by the hypotheses and conceptual model. Then, we outline the methodology of the study as well as the operationalization of the modelling techniques. Next, we present the study results followed by the conclusion and discussion. Finally, we present the limitations and recommendations for future research.

2. Theoretical Framework

To understand the relevancy of customer value and customer churn for firms, we first review the impact of database marketing on the way firms serve individual customers. The customer lifetime value as a metric for individual customer value is then described. The dependent variable of customer churn is then established as a crucial component of customer lifetime value. We dive into the identified limitation of customer churn predictors and the role of customer churn intent, on which the conceptual research model is based.

2.1. Database Marketing

advantage for firms (Leeflang, Verhoef, Dahlström & Freundt, 2014). In this information-intensive business climate, database marketing has emerged as an invaluable approach for achieving such added value. Database marketing can be defined as using customer databases to enhance marketing productivity through more effective acquisition, retention, and development of customers (Blattberg et al., 2008). Through database marketing, firms can identify for which individual customers' marketing efforts will pay off, allowing for personalized targeting efforts. By doing so, a firm can more effectively use marketing resources and increase marketing productivity.

For database marketing to be a successful strategy, the collection of data is of paramount importance. In this context, an organization needs to have data on individual customers. The need to have data on individual customers has strengthened the need for firms to structure the organization around individual customers' needs. The philosophy of marketing decision-making that is based on serving customers in a long-term relationship is known as a customer-centric strategy (shah et al. 2006). For companies that possess individual customer data, customer-centricity has been the prevailing management philosophy since the 90’s of the last century.

2.2. Customer lifetime

In a customer-centric organization, the goal of customer management is to create long-term value of customers in order to increase the customer lifetime value (CLV). The customer lifetime value is the overall value a firm derives from its customer. The idea of the CLV is to capture the value of a customer over the entire past and expected length of the relationship between a customer and a firm. Thus, a significant part of customer management has been managing and increasing the customer lifetime metric (Kumar & Reinartz, 2012).

A key driver of customer lifetime value is customer churn. Customer churn refers to the termination of the contractual or non-contractual relationship between the customer and a firm (Leeflang, Wieringa, Bijmolt & Pauwels, 2015). For instance, when a customer decides to terminate a subscription to a video-streaming service, a customer churn event is recorded. Customer churn is part of the retention part of the CLV, as when a customer churns it impacts the length of the relationship between a customer and firm.

agreements, immediate sales are lost, and future cash flow through cross- or upselling (Gupta, Lehmann, and Stuart 2004). Customer churn rates highly differ across industries and directly impact the value of a firm. It is estimated that a churn rate of over 20 percent in the

telecommunications industry can reduce a firm’s value by several billion dollars (Gupta, Lehmann & Stuart, 2004). In this particular industry, annual churn rates of below 20 percent are an exception (Blattberg et al., 2008). Winer (2001) reports that retention is a far more powerful tool for internet firms' market value than acquisition. Overall, the benefits of a long-term relationship outweigh the acquisition efforts.

For firms that have a contractual relationship with their customers (e.g. video-streaming services), customer churn is characterized by the termination of a contractual agreement between the customer and a firm. In a contractual setting, the recorded moment of customer churn is determined when a client interrupts their relationship with the firm (Buckinx et al., 2005). A substantial body of literature on retention management has focused on churn in a contractual setting. Customer databases from the telecom industry, banking sector, and insurance services have commonly been used for research that seeks to find predictors of customer defection (Risselada, Verhoef & Bijmolt 2010; Lemmens, Croux, 2006; Holtrop, Wieringa, Gijsenberg & Verhoef, 2017).

In a non-contractual setting of churn, the moment a customer interrupts a firm's relationship is not immediately apparent, as customers can continuously switch between competitors without customer-to-firm interactions. In this context, firms often lack insight into a customer's switching decisions at a certain point in time (Leeflang et al., 2015). In both contexts, customers' switching behavior is very common (Dwyer, 1997), which makes understanding the drivers of such behavior highly valuable to a firm.

2.3. Drivers of churn

Given the potential problems of losing customers, it is essential for companies to first identify the drivers that cause or prevent customer churn. We elaborate on predictors of customer churn that have been established in previous literature. Next, we make hypotheses based on the limitations that we identify of a set of these predictors.

number of previously researched factors. Whether it is worth it to stay, depends not only on ‘financial’ cost of switching. Next to the financial costs of switching, procedural costs and relational costs also play a role (Burnham, Frels & Mahajan, 2003). The financial costs are directly related to the monetary or benefit loss that is incurred by switching, while the procedural costs refer to the expenditure, time and effort a consumer has to consider. The relational cost pertains to the psychological or emotional discomfort a customer experiences due to the loss of the relationship of leaving a firm (Burnham, Frels & Mahajan, 2003). Previous literature has shown that the focus on price leads to a reduced customer loyalty (Ansari, Mela, & Neslin, 2008; Sinha, 2000)

Customer satisfaction. Another key determinant of customer churn is customer satisfaction. Previous literature by Rust and Zahorik (1993) established a link between customer satisfaction and retention. In retention management, customer satisfaction is a commonly used customer feedback metric as a predictor of customer churn (de Haan, Verhoef & Wiesel, 2015). In essence, customer satisfaction is a psychological notion by an individual that relates to the extent to which the individual's needs are met. As such, we define customer satisfaction as the emotional evaluation (Song, Wang, & Han, 2019) regarding price to quality ratio, service to quality ratio, fit-to-needs and meeting expectation (Blattberg, Kim and Neslin, 2010). From the formal definition, it can already be implied that customer satisfaction is not a unitary construct. Instead, we take a broader definition of customer satisfaction, and define customer satisfaction as a combination of both an economic and service-related construct (Geyskens, Steenkamp & Kumar, 1999). The economic determinant of satisfaction pertains to the customers’ financial outcomes, while the service determinant pertains to the level of service a customer experiences (Geyskens, Steenkamp & Kumar, 1999; Wirtz et al., 2014).

Customer satisfaction scores are a commonly used method by firms to gain customer insights. The underlying idea is that customers will increasingly be satisfied with the product/service by improving the aspects mentioned above. Presumably, more satisfied customers experience higher value from the products/services they use, leading to increased customer loyalty and retention.

can be made in predicting customer switching behavior. For instance, Mittal and Kamakura (2000) found that somewhat dissatisfied and completely dissatisfied customers were equally unlikely to churn. In a linear approach, it was assumed that a dissatisfied customer was less likely to churn than completely dissatisfied customers

Dissatisfied customers experience a negative emotional evaluation of the extent to which an individual’s needs are met. When customers are dissatisfied, they are increasingly motivated to search for alternative choices (blattberg et al., 2008; Anderson & Mittal, 2000). Subsequently, stated dissatisfaction can act as an important predictor of switching behavior.

We hypothesize that the economic and service determinants of customer satisfaction can be used to predict customer churn of dissatisfied customers. Thus, we deem the economic and service determinants of satisfaction as a potential driver of customer churn that are relevant for our research. The economic determinants focus on the level of pricing and economic satisfaction. Given this classification, we propose the following hypothesis:

H1a: The likelihood of customer churn increases when a customer experiences dissatisfaction in costs.

Moreover, we classify satisfaction on the service determinants as the quality of service a customer receives. In prior research, a firm's service quality is defined as the overall

impression a customer has of the superiority in the services a firm offers relative to switching options (Bansal & Taylor, 1999). Given this classification, we propose the following hypothesis:

H1b: The likelihood of customer churn increases when a customer experiences dissatisfaction in service quality.

It should be noted, however, that not every customer experiences pricing and quality levels the same way. A study conducted by Mittal and Kamakura (2001) finds that customer satisfaction is not generalizable across all customers, but that customer characteristics as age, education and income can significantly affect the probability of repurchase behavior by a customer. We propose that customer characteristics can act as a relevant driver of customer churn. To test whether customer characteristics influence the likelihood of a customer churning, the following hypothesis has been created.

Next to the economic and service-related determinants of satisfaction, we propose customer engagement to be a relevant customer churn predictor. Kumar, Aksoy, Donkers, Venkatesan, Wiesel & Tillmanns (2010) argue that the transaction-based customer lifetime value is not sufficient to capture the indirect value customers provide to a firm. It is argued that individual customers' value is determined not only by the amount of individual revenue they generate during the customer-firm relationship but also by the engagement of a customer with a firm (Kumar et al., 2010).

Customer engagement. Customer engagement can be defined as behavioral manifestations by customers in a customer-to-firm relationship that go beyond purchase, resulting from motivational drivers (Van Doorn et al., 2010). For firms it is interesting to measure and stimulate customer engagement. It allows them to value their customer base by analyzing who their most engaged customers are. Doing so allows a firm to segment its customer base by classifying customers as having a high-engagement or low-engagement with the firm. For instance, video-streaming service ‘Netflix’ now has a forum where users can review movies. This provides Netflix with rich data on their offerings, but also with customer specific engagement data that can be incorporated in predictive modelling methods. In customer-to-customer relationships, customer-initiated engagement can consist of offline Word-of-Mouth (WOM), online customers helping other customers, sharing and spreading brand-related information or even co-creation through suggestions for product and service improvements. Between the customer and the firm, firms can enable customers to acquire and provide information (e.g. online chat, personal interactions).

Firm-initiated customer engagement is deliberately initiated by marketers to control and exploit the customers' readiness to engage. This includes voting behavior in firm-initiated contests, co-creation, and brand communities (Kumar et al., 2010). Other ways of measuring customer engagement are via the extent to which customers engage in loyalty/reward programs, user-reviews and star ratings (Doorn et al., 2010).

We argue that customer engagement has a negative impact on customer churn, as engaged customers experience a stronger relationship with the firm than passive customers. Prior research has established that customers' commitment to a brand or firm positively affects customer retention (Verhoef., 2003). In line with this reasoning, we propose the following hypothesis:

H2: The likelihood of customer churn decreases when a customer engages with the current product/service provider.

Commented [dd1]: If positive, conclude that it is relevant

2.4. Churn intent

A challenge for firms in predicting customer churn for retention management is that it is mostly focused on classifying customers who churn and a customer who do not churn (de Vriendt, Berrevoets & Verbeke, 2020). However, it can be interesting for firms to have insight into customers who have an intention to switch but are still susceptible to retention campaigns by the current firm. To distinguish between definitive churners and customer who intend to churn allows marketing managers to allocate retention management resources better.

Whereas customer churn is characterized by the termination of a contractual or non-contractual relationship between a customer and a firm (Leeflang et al., 2015), customer churn intent does not capture the actual termination. Thus, we define customer churn intent as the self-reported likelihood of terminating a current relationship with a firm (Wirtz et al., 2014). The difference between customer churn and customer intent to churn is that the intent to churn captures a hypothetical situation. We can assume that in this scenario, the customer has a greater chance of being retained. This makes the customer more interesting to approach for retention efforts.

Prior research has found that churn intent can directly affect the likelihood of a customer actually churning (Bansal & Taylor, 1999). This research finds that in the context of telecom providers, the stated intent to churn has a direct positive relationship with the actual switching behavior of the customer. Thus, we suspect that churn intent may act as a mediator on actual customer churn in our research. We propose the following hypothesis:

H3a: The likelihood of customer churn is mediated by the intention to churn.

evaluating alternatives) are the central aspect of the decision-making process (Liberman et al., 2007).

As we distinguish between customer churn likelihood and customer intent to churn, we suspect that non-monetary switching costs can predict churn intent. We propose the following hypothesis:

H3b: The likelihood of a customer expressing an intent to churn increases when a customer experiences dissatisfaction in service quality.

Lastly, we reason that customers that experience a stronger relationship with the firm are less likely to switch to a competitor. In line with this reasoning, we hypothesize that this

relationship also holds for the intention to switch to a competitor. We propose the following hypothesis:

H3c: The likelihood of a customer expressing an intent to churn decreases when a customer engages with the current product/service provider.

2.5. Predictive modelling

In retention management, predictive modelling is a common form of analysis conducted by firms with customer databases. Through predictive modelling using statistical methods, customer behavior can be modeled and predicted. For instance, the susceptibility to personalized offerings or likelihood t to churn given a set of input variables are of particular interest (Blattberg, Kim & Neslin, 2008).

either marginal differences across modelling methods, or modelling performance differences that are attributable to the type of data that used (Risselada, Verhoef & Bijmolt, 2010). A study by buckinx et al. (2005) in a non-contractual churn setting find no significant

differences in model performance across modelling techniques. In this research, we compare logistic regression, classification trees, support vector machines, boosting, bagging, and random forest as a method to predict churn to discover differences in model performance.

2.6. Conceptual Model

Based on the literature review, we conceptualized the hypothesized relationships in a

conceptual model (Figure 1). The model captures the hypothesized direct and mediated effects of the predictor variables on intended customer churn and actual customer churn. In line with our research at hand, the model is limited to the predictors that we expand on in the

established literature.

3. Methodology

This chapter discusses the dataset that was used, the variables, and the statistical methods to test the stated hypotheses. In the first part, we discuss how the data is collected and prepared. Next, we elaborate on the specific variables the dataset contains that have allowed us to conduct predictive modelling. In the second part, we specify, estimate, and compare the predictive models.

3.1. Data Description

For this study, we excluded non-relevant variables, merged datasets, restructured the dataset and transformed the data into variables suited for predictive modelling.

3.1.1. Data collection

A panel research company collected the data we used to conduct this research from October 2013 until January 2014. The dataset contains customer level information across 63 competitors on the Dutch health insurance market. It contains the individual customer responses to a questionnaire about health insurances. Customers of a health insurance company were asked what made them consider switching to a different health insurance company. Moreover, the dataset contains the actual observed switching behavior. In addition to the stated and observed switching behavior, customers have been asked what aspects play a role in their decision making when purchasing a health insurance plan. Next, the way the customer has acquired information in both on-and offline settings has been recorded. Lastly, customer demographics have been recorded in a separate dataset.

3.1.2. Data preparation

The dataset consists of customer-level data in which each customer has up to 4 points in time in which their behavior is observed. The data was observed in a one-to-many fashion, meaning that each observation of an individual customer is expressed as an individual row in the data over multiple observations.

First, we merged the primary dataset with the demographic’s dataset so that every customer row is now accompanied by background information that is specified in the variable specification section.

variable values (e.g., churn across time points) are given a separate column. Thus, this long to wide restructuring has led to a decrease in the number of rows but an increase in number of variables in the dataset.

The aforementioned procedure leaves us with 6645 observations of individual customers with 18 relevant variables. In subsequent analysis, we used the programming language R.

3.1.3. Variable explanation

The type of buyer variable in the dataset can be used to classify churners. The type of buyer variable classified customer purchasing behavior as either ‘passive extender,’ ‘active extender,’ and ‘switcher’. To be able to represent churn in a binary fashion, we transformed this variable so that all switchers are grouped as churners and the remaining observations as non-churners.

Next, customers were asked for the type of search they conducted in acquiring information on health insurance plans. This could be in an online environment and in an offline environment. Table 1 visualizes the different ways customers indicated they acquired information on health insurance plans.

Table 1 Information variables Information acquired Method

Online

Visited website E-mail Advisor Online chat with Advisor Social Media (e.g., Facebook, twitter)

Offline

Conduct family and/or friends (WOM) Call Advisor

Personal chat with Advisor Read brochures and magazines

Read insurance offerings TV, Radio

Based on the responses visualized in table 1, we identified the variables that act as proxy of customer engagement. We created two dummy variables. The customer engagement variable consists of ‘online chat/call/mail with an advisor’, ‘Social Media’, Conducting family and/or friends (WOM)’. The non-engagement variable consists of ‘Visited website’, ‘Read

Commented [JW2]: Je data is veel rijker dan je theorie

doet vermoeden. De theorie kan nog wat dieper denk ik, met mogelijk aanvullende hypotheses. Bijv. modererende effecten van online/offline, of modererende effecten van satisfaction determinant

brochure/insurance offerings’ and ‘TV & Radio’. Customers may have used multiple information acquisition methods. Hence, we include both the engagement and non-engagement dummy variables in the model.

As part of the data collection, customers were asked what aspect of satisfaction would lead them to switch to another health insurance plan. The aspects that they deemed important in making a switching or staying decision for a health insurance plan are shown in table 2. These aspects were answered on a scale of 1 to 4.

Table 2 Satisfaction variables

Variable Aspect Satisfaction Determinant SX7A Price of current health

insurance price too high

Economic SX7B Unreasonable price increase

of current insurance SX7C Lower price at different

health insurer

SX7G Attractive offer different health insurer

SX7F

Attractive collective discount at different health insurer

SX7H Difficulties with paying current price

SX7D Dissatisfied with coverage of current health insurance

Service SX7E Better service coverage at

different health insurer SX7K Dissatisfied with current _{insurer’s contractual policies} SX7M Dissatisfied with service of

current health insurer

SX7I Changes in personal

situation

Other SX7J _{Changes in personal health}

SX7L _{Political discussions}

In addition to actual observed churn behavior, customers were asked to what extent they intent to churn across the measurement moments. The intent to churn could be indicated as ‘yes, for sure’, ‘maybe’ and ‘no intent to switch’. Over the course of time, customers may have indicated a varying score based on their perception at that moment. For instance, a customer might indicate no intention to churn at the first observation but indicated a clear intention to churn at the last observation. We transformed the given scores of the measured moments into a dummy variable that indicates either no intent to churn or intent to churn. The intent to churn is defined as a person who has indicated ‘yes, for sure’ at least once across the observations. Using a strict dummy that classifies definitive churn intentions provides us with confidence that the intent to churn of a customer has been evident at some point in the observations, which benefits the predictive modelling power.

Next, we transformed the demographical data into more intuitive categories. Table 3

Demographic variables

Variable Category Recoded Category

Age Ranging from 18 - 113

18 - 35 36 - 55 56 - 75 76 - 113

Education 14 categories ranging from ‘no education’ to ‘WO-doctorate’ No education Special Education Low education Medium education High education Income

20 categories ranging from ‘no income’ to ‘€4100 or more’

No income = € 0 Low income = up to €1700 Average income = €1701 - €3400

High income = €3401 +

distribution of age of respondents is right-skewed (𝜇 = 52.63). Thus, we classified the age to represent the higher age better. Then, we recoded the educational categories to better represent no, special, low, medium, and high education (CBS, 2020). Special Education represents education for people with a handicap that requires additional care at school. Low level education represents elementary education. Middle level education represents applied education (MBO). High level education represents professional education (HBO, WO). Then, we recoded income to represent intuitive levels. Given the average net income in The Netherlands of €2400, we recoded the income levels to represent a low, medium and high income as presented in table 3.

3.1.4. Outliers and missing values

Next, we checked the combined and transformed dataset for missing data and oddities. When looking at the total number of observations, we find that out of 6445 customers a number of 744 customers churn, and 4612 customers do not churn. This leaves us with 1089 customers of whom it is unclear whether they switched or not. These cases are left out of the churn analysis. Out of the 6645 observations, we find 4022 observations in which the customer did not indicate how their information search took place. As this information is crucial for our theorized customer engagement predictor of customer churn, we exclude these cases from the analysis. Moreover, out of the 6445 observations, we find 35 cases in which the respondent’s churn intent was not recorded. Upon closer analysis these cases appear incomplete and have thus been excluded from analysis.

Regarding the customer demographics, out of the 6445 we find 531 cases in which the customer's income demographic has not been stated. Income being a theorized predictor of customer churn within the customer background information, we have excluded these cases from the analysis.

The final dataset suited for analysis contains 1939 observations of which 293 observed churners.

3.2. Model specification

We use a binomial logistic regression to predict the dependent variable of customer churn. Then, we use conventional and machine learning methods for classifying churners and non-churners in an out-of-sample manner. Additionally, we conduct a mediation analysis.

Commented [JW4]: Wow. Dat is nogal een aderlating!

Commented [JW5]: Het is niet zo duidelijk waarom je de

verschillende modellen gebruikt voor het schatten van churn.

3.2.1. Binomial Logistic Regression

First, we use a binomial logistic regression to predict the switching behavior of a customer. Binomial logistic regression is a binary choice model. In the current research, customer churn is a binary variable in which 1 indicates a customer churning and 0 indicates no-churn. When a customer has switched to another health insurance, a churn event takes place. We estimate whether customer i given period t churns (0/1) given the predictor variables xi. An unobserved utility value characterizes a binary logistic model. This latent utility is a value on an arbitrary scale that is transformed to a probability value between 0 and 1. Using the transformed value, the choice of the customer is determined. To transform the utility into a probability, we define a latent utility function. The logit model is specified as follows (Fok, 2017):

𝑈𝑖= 𝛼 + 𝑥𝑖′𝛽 + ε𝑖

where xi is the vector of predictors of the dependent variable, is the intercept, is 𝛼 vector of coefficients and an error term ε𝑖. Next, we link the latent utility 𝑈𝑖 to the outcome probability as follows (Fok, 2017):

𝑌𝑖= {

0 𝑖𝑓 𝑈𝑖≤ 0 1 𝑖𝑓 𝑈𝑖> 0

We assume that the error term ε𝑖 is independently distributed according to the cumulative distribution function (CDF) 𝐹(∙). Thus, the probability of churning (𝑌𝑖= 1) is:

𝑃[𝑌𝑖= 1] = 𝑃[𝑈𝑖> 0] = 𝑃[𝛼 +𝑥𝑖′𝛽 + ε𝑖> 0] = 𝑃[ε𝑖> −𝛼 −𝑥𝑖′𝛽] = 1 − 𝐹(−𝛼 −𝑥𝑖′𝛽) = 𝐹(𝛼 +𝑥𝑖′𝛽)

We use the CDF to determine the probability of a random observation in the data. Using the CDF, we write the probability of a customer churning as:

𝑃[𝑌𝑖= 1] = Λ(𝛼 +𝑥𝑖′𝛽)=

exp (𝛼 +𝑥𝑖′𝛽) 1 + exp (𝛼 +𝑥𝑖′𝛽)

3.2.2. Modelling Methods

To compare the predictive performance of binomial logistic regression with other modelling techniques, we use tree classification, bootstrap aggregation, boosting, random forest, and a support vector machine.

Tree Classification. First, we use a classification tree to classify churners and non-churners. To build the tree model we split the entire dataset into mutually exclusive discrete subsets so that each customer is assigned to a subset. The goal of the tree is to predict the dependent variable of churn based on the values or attributes of the independent variables (Leeflang, Wieringa, Bijmolt & Pauwels, 2017). In other words, we build a classification tree to assign a class for each observation with the least number of splits. The tree consists of 3 type of nodes: root, children and terminal. Starting from the root node, the algorithm attempts to continuously split children nodes using its own criterion function to find the best split. The tree classification algorithm ends at a terminal node, when the splitting value is not able to exceed the pre-specified stopping threshold any longer. At each split a particular predictor variable for predicting churn forms the basis for the decision rule that drives the split. The decision rule follows the best available strategy by selecting the variable and threshold of classification are based on the best improvement. By default, a tree model accommodates for interaction effects and nonlinearities (Blattberg, Kim and Neslin, 2010).

Bootstrap Aggregation. Then, we run a bootstrap aggregation algorithm to establish the model performance. The goal of the bootstrapping algorithm is to train a decision tree so that it can be used to make out-of-sample predictions. Bootstrap aggregation, or ‘bagging’ is a method in which new dataset are bootstrapped with an equal number of observations as the original dataset. These datasets are based on random draws from the observations of the original dataset. Bootstrapped datasets can contain the same observation across multiple bootstraps. However, the complete bootstrapped datasets will always differ from one another (Leeflang et al.,2017). The predictions that are made aggregated over all generated decision trees so that an average final prediction can be made. This method of aggregation leads to well-generalized models as overfitting is controlled for.

misclassifications of conversations. Boosting is a flexible method as it allows for outliers and is suitable for data that contains a lot of missing.

Random Forest. The random forests machine learning algorithm works much like bagging. However, the fundamental difference between bagging and random forests is that random forests only use a randomly selected subset predictor variables to determine the tree split. This leads to every variable has the opportunity to generate a split in the prediction tree. When all variables are used to generate tree splits, relative variable importance can be deduced. This can then be used to make inferences about the importance of each individual predictor.

Support Vector Machine. Another frequently used classification technique is the Support Vector Machine (SVM). Again, a training and evaluation dataset is used for the algorithm to train for out-of-sample predictions. A Support Vector Machine algorithm splits observation into separate classes based on the features of the training dataset. A linear approach is used to separate the various observations into homogeneous classes. However, as a straight line is often not able to strictly separate the observations into one class or the other, the goal is to achieve a balance between model fit and model validity (Leeflang et al., 2017).

3.2.3. Mediation analysis

To test the mediation hypothesis, we formulate a mediation model to test if churn intent acts as a mediator to actual customer churn. Mediation is a process through which we assume that there is a causal relationship between the dependent and independent variables. In our case, we test whether the actual observed churn behavior results from customers' intentions to churn. The mediation process can be defined as complete mediation and partial mediation. In complete mediation, the independent variable (𝑥1𝑡) no longer affects the dependent variable (𝑦1𝑡) once the mediation variable (𝑥2𝑡) is controlled for.

Figure 2

Mediation Schematical representation

We follow the mediation steps, as outlined by Baron and Kenny (1986). First, the effect of the independent variable on the dependent variable is established (path c). Then, the independent variable must affect the mediator (path a). We test for the effect of the mediator on the dependent variable (path b). When controlling for the effect of the mediator, the independent variable no longer or only partially affects the dependent variable (path c’).

3.3. Model estimation

A logistic model does not require testing many of the assumptions that apply to a General Linear Model (GLM) or linear regression models (Leeflang, Wieringa, Bijmolt & Pauwels, 2017). For instance, the assumptions regarding linearity, normality and homoscedasticity are eased. However, some assumptions still apply.

To start, we have set the dependent variable to be a binary churn variable (0/1) required for estimating a logistic model. We need a sufficient sample size and take 10 cases for each independent variable at a minimum as a guideline. We conclude that with 4458 observations, we have a sufficient number of cases for logistic modelling given the number of modelling variables.

reference category. Across the economic and service aspects, the reference category is set as ‘not important’, meaning that the following categories represent an increase in importance up to ‘highly important’. For the variables age, income and education we take the first category as the reference level, being ‘18-35’, ‘no education’ and ‘no income’, respectively.

Next, we check whether the multicollinearity assumption holds across the predictors. We conduct a VIF score test. In this research, we take a higher VIF score than 5 to be a

multicollinearity issue (Leeflang et al. 2015).

To deal with the trade-off between goodness of fit and model simplicity, we use the Akaike Information Criterion (AIC). This procedure selects the best model based on the AIC, i.e., the model with the lowest AIC value. Variables are excluded in a stepwise manner to obtain the most parsimonious model with the lowest AIC value. In addition, we use the likelihood ratio test (LR), Cox & Snell pseudo R2, Nagerkerke pseudo R2, and McFadden pseudo R2. Based on the above measures, we choose our preferred model.

The formulation of the full model in relation to observed churn behavior is as follows: 𝑈𝑖= 𝛼 + 𝛽1𝑎𝑔𝑒𝑖+ 𝛽2𝑖𝑛𝑐𝑜𝑚𝑒𝑖 + 𝛽3𝑒𝑑𝑢𝑐𝑎𝑡𝑖𝑜𝑛𝑖 + 𝛽4𝑒𝑛𝑔𝑎𝑔𝑒𝑚𝑒𝑛𝑡𝑖 + 𝛽5𝑛𝑜𝑛_𝑒𝑛𝑒𝑔𝑎𝑔𝑒𝑚𝑒𝑛𝑡𝑖+ 𝛽6𝑆𝑋7𝐴𝑖+ 𝛽7𝑆𝑋7𝐵𝑖 + 𝛽8𝑆𝑋7𝐶𝑖+ 𝛽9𝑆𝑋7𝐷𝑖+ 𝛽10𝑆𝑋7𝐸𝑖+ 𝛽11𝑆𝑋7𝐹𝑖 + 𝛽12𝑆𝑋7𝐺𝑖+ 𝛽13𝑆𝑋7𝐻𝑖+ 𝛽14𝑆𝑋7𝐼𝑖+ 𝛽15𝑆𝑋7𝐽𝑖 + 𝛽16𝑆𝑋7𝐾𝑖+ 𝛽17𝑆𝑋7𝐿𝑖+ 𝛽17𝑆𝑋7𝑀𝑖 (1) 3.4. Model Comparison

4. Results

This section describes the results of our study. We first provide a sample description to gain an insight in the nature of the observations. Then, we test our hypotheses. In order to test our hypotheses, we conducted logistic regression analyses. In addition, we conduct out-of-sample predictive analysis. We validate, estimate, and compare the predictive models and draw conclusions.

4.1. Description of the sample

We start with a description of the obtained sample using demographic variables. These include income, education, and age (figure 3, figure 4, figure 5). For income levels, the largest group of respondents are in the average income category. With regards to age, we find

Figure 3 Education Levels

it notable that the sample is skewed towards higher age. Compared to the general Dutch population the special education and high education categories are slightly overrepresented in the sample (Onderwijsincijfers, 2019). This might be explained by the increased need for good healthcare related to special education, or the increased awareness of switching options for higher education.

Figure 5 Age Categories

Next, we find the number of customers who churn and the number of customers who intent to churn, next to customers who have not churned or intended to churn (figure 6, figure 7) in the remaining observations.

4.2. Predictive modelling of churn behavior

Now estimate the observed churn behavior to test hypothesis H1 and H2. First, we validate the full model, then we proceed with the most parsimonious model.

4.2.1. Validation Churn Model

We first check for the assumption of no multicollinearity between predictor variables of model (1). As we find all values to be below 5, we conclude that the assumption of no multicollinearity between variables holds (table 4).

Table 4

VIF Scores of Churn Predictors VIF Df Age 1.215 3 Income 1.158 2 Education 1.234 3 Engagement 1.031 1 Non-Engagement 1.041 1 SX7A 3.218 3 SX7B 2.842 3 SX7C 3.396 3 SX7D 3.351 3 SX7E 2.842 3 SX7F 2.035 3 SX7G 2.878 3 SX7H 2.253 3 SX7I 2.552 3 Figure 7 Customer Churn Intent Figure 6

SX7J 2.584 3

SX7K 2.749 3

SX7L 1.587 3

SX7M 3.131 3

Next, we aimed to obtain the most parsimonious model. We conduct an Akaike Information Criteria (AIC) procedure to obtain the model. This model selection procedure is characterized by punishing the addition of parameters that do not increase the model's overall fit. As such, the AIC deals with the risks of overfitting and underfitting. We find the most parsimonious model with the lowest AIC value (model 2).

Next, we conduct a likelihood ratio test to assess the parsimonious model's goodness of fit compared to the null model. We conclude that the parsimonious model is performing significantly better than the null model. Nevertheless, model (2) does not perform

significantly better than model (1). Lastly, we compare the Pseudo R2 _{scores of both models.} We find a marginal difference in Pseudo R2 _{values, with model 1 performing better. This is} expected, as the higher number of variables leads to increased model noise and worse predictive accuracy. Based on the validation measures (table 5), we proceed with model (2).

Table 5

Model Validation Measures

AIC Likelihood

Ratio Test Pseudo R2

P Value CoxSnell Nagelkerke McFadden

Null Model 1648 0 0 0

Model 1 1616 0.8289 0.0651 0.1138 0.0793

Model 2 1580 < 0.01 0.0550 0.0961 0.0666

4.2.2. Logistic regression

In order to test hypotheses H1 and H2, we now estimate and interpret model (2). The coefficients of the logistic regression are found in appendix A. We start by looking at the signs of the coefficients, as in logistic regression, the coefficients have a non-linear

relationship with the probability of churn. The estimates are based on the latent variable of the regression equation. Thus, the model estimates are only used to determine the sign of the

𝑈𝑖= 𝛼 + 𝛽1𝑎𝑔𝑒𝑖+ 𝛽2𝑆𝑋7𝐵𝑖+ 𝛽3𝑆𝑋7𝐶𝑖+ 𝛽4𝑆𝑋7𝐺𝑖+ 𝛽5𝑆𝑋7𝐼𝑖 + 𝛽6𝑆𝑋7𝐾𝑖+ 𝛽7𝑆𝑋7𝑀𝑖

coefficients in which a positive sign indicates an increase in the probability of churn and a negative sign indicates a decrease in the probability of churn.

Coefficient Signs. We find both negative and positive effects across economic and service determinants of satisfaction. For the economic determinants of satisfaction, we find a significant positive coefficient sign for aspect C. As hypothesized in H1a this indicates that dissatisfaction in cost a churn option has a positive effect on churn likelihood. Contrary to H1a, we find negative coefficient signs economic determinants B and G. Thus, we can only partially confirm hypothesis H1a.

For the service determinants of satisfaction, we find a significant positive coefficient sign for aspect M. As expected, this indicates that dissatisfaction of a customer with the current service that is offered by the Health insurance provider leads to an increase in customer churn. Contrary to H1b, we find a significant negative coefficient sign for service-related determinant K. Thus, we can only partially confirm hypothesis H1b. Next, we find a negative effect of age categories on churn probability, indicating that being in a higher age group decreases customer churn likelihood. While education and income do not play a significant role in predicting customer churn, we can partially confirm H1c regarding the effect of age on customer churn likelihood.

Lastly, we find that aspect I has a significant positive coefficient. This indicates that changes in the personal situation of a customer has a positive effect on customer churn. In the estimation of model (2), the hypothesized customer engagement variables do not significantly affect the likelihood of a customer churning.

Marginal Effects. To shed more light on the effect size of the estimated coefficients, we look at the marginal effects. We use the marginal effects to determine the change in probability of observing churn over no-churn when the predictor variables change from 0 to 1. As we deal with categories, we find marginal effects relative to the reference category (table 6). We only interpret significant results and relative to the reference category of ‘not important’.

As for service-related determinants of satisfaction, we can say that a marginal increase in dissatisfaction with the current service leads to an increase of customer churn by 7 percentage points. Also, a marginal increase in dissatisfaction with the current contract has a negative effect on the probability of a customer churning by 3.8 and 5.5 percentage points relative to the reference category.

For the age categories, we find that a marginal increase in the age category leads to a decrease in churn probability of 4.4, 8.1, and 11.8 percentage points, respectively. This indicates that the higher the age category, the higher the probability of churning. Lastly, we find that a marginal increase in personal situation leads to an increase of 6.5 percentage points in churn probability relative to the reference category.

Table 6

Marginal Effects Customer Churn

Category Variable Marginal Effect SE P value Age Age 2 -0.0446 0.0190 <0.05 Age 3 -0.0812 0.0195 <0.01 Age 4 -0.1186 0.0169 <0.01 Economic SX7B 2 -0.0473 0.0185 <0.05 SX7B 3 -0.0533 0.0191 <0.01 SX7B 4 -0.0783 0.0182 <0.01 SX7C 2 0.0298 0.0378 0.42 SX7C 3 0.0922 0.0360 <0.05 SX7C 4 0.1600 0.0468 <0.01 SX7G 2 -0.0439 0.0217 <0.05 SX7G 3 -0.0221 0.0230 0.33 SX7G 4 0.0251 0.0316 0.42 Service SX7K 2 -0.0288 0.0194 0.13 SX7K 3 -0.0370 0.0205 0.06 SX7K 4 -0.0550 0.0214 <0.01 SX7M 2 0.0774 0.0286 <0.05 SX7M 3 0.0234 0.0273 0.39 SX7M 4 0.0544 0.0383 0.15 Other SX7I 2 0.0650 0.0268 <0.05 SX7I 3 0.0006 0.0216 0.97 SX7I 4 0.0419 0.0308 0.17

Categories: 2 = ‘slightly important’, 3 = ‘important’ 4 = ‘highly important’, age 2 = 36-55, age 3 = 56-75 age 4 = 76-113, reference categories = ‘not important’, age 1 = 18-35

does not affect the odds of customer churn, given that all other independent variables stay constant. We only interpret significant results.

As for economic determinants of satisfaction, we find that if the importance of perceived price at a competitor increases by one unit, the odds of a customer churning increase by a factor of 1.28, 2.13 and 3.19 as compared to the reference category. A one-unit increase in the perceived price of the current health insurer decreases the odds of a customer churning by a factor of 0.63, 0.60 and 0.43 as compared to the reference category. Moreover, a one-unit increase in attractiveness of offers at competitors decrease the odds of a customer churning by a factor of 0.65 as compared to the reference category.

As for service-related determinants of satisfaction, we find that if the dissatisfaction with the current service increases by one unit, the odds of a customer churning increase by a factor of 1.80. A one-unit increase in dissatisfaction with the current contract leads to a decrease in the odds of a customer churning by a factor of 0.69 and 0.56 as compared to the reference category.

We find that if the age increases, the odds of a customer churning decrease by a factor of 0.21, 0.66 and 0.47 as compared to the reference category.

Lastly, we find that a one-unit increase in importance given to personal situation change increases the odds of a customer churning by a factor of 1.65.

4.3. Model performance comparison

For the purpose of predictive forecasting, we now assess the model performance across machine learning techniques. The logit model is compared to a classification tree, support vector machine, random forest, boosting, and bagging methods. To enable out-of-sample forecast, we first split the dataset into a training and evaluation set. The training set contains 25 percent of the sample that was used for predicting actual churn behavior. The evaluation set contains the remaining 75% This means that the training dataset now contains 484 observations, and the evaluation dataset 1455 observations.

Logistic Regression. We find that the logit model can classify churners and non-churners correctly in 68.59 percent of the time. Figure 8 shows the lift plot, which indicates an improved performance over the null model. The TDL shows that our model performs 1.94 times better than the null model when classifying the top 10 percent of actual churners.

The Gini coefficient of 0.32 also shows that our model performs slightly better than a random model in predicting customer churn.

Commented [JW7]: Logistic regression is the best?

Figure 8

Logistic Regression Lift Curve

Tree classification. Next, we use a tree model to classify churners and non-churners. Using the tree model, we find that churners and non-churners are classified correctly 69.07 percent of the time. The TDL indicates an increase in the predictive performance of the top 10 percent churners of 1.63 times. The Gini coefficient of 0.14 is a slight improvement over the null model (appendix C).

Bagging. Next, we perform the bootstrap aggregation or ‘bagging’ to classify churners and non-churners. We find that 83.43 of the churners and non-churner are correctly classified using this method. We find a TDL of 1.28, indicating the model performs 1.28 times better than the null model when classifying the top 10 percent churners. Lastly, we find a Gini coefficient of 0.17, indicating a slight performance increase over the null model (appendix C).

Boosting. Following, we use a boosting method to classify the churners and non-churners. Using the boosting method, we find that churners and non-churners are classified correctly 78.62 percent of the time. With a TDL of 1.41 and a Gini coefficient of 0.21, the boosting model performs marginally better than the null model (appendix C).

Random Forest. Next, we use a random decision forest algorithm to classify churners and non-churners. We find that 82.88 percent of the time, the churners and non-churners correctly classified. We find that the random forest classification performs 1.58 times better at classifying the top 10 percent churners than the null model with a Gini coefficient of 0.32. This is an improvement over the null model (appendix C).

84.32 percent of the time. The SVM performs 1.58 times better at predicting the top 10 percent churners than the null model with an overall Gini coefficient of 0.1. This is slightly better than the null model (appendix C).

4.3.1. Model Performance Conclusion

Concludingly, we find differences in model performance measures across classification algorithms. The most commonly used logistic regression model performs best in classifying the top 10 percent churners with a TDL of 1.94. The support vector machine performs best when classifying churners and non-churners correctly, with a hit-rate of 84.32 percent. Overall, the Gini coefficient indicates a slight increase in predictive performance over the null model across all classification algorithms.

4.4. Mediation analysis

To test the hypothesized effect of churn intentions acting as a mediator on actual churn, we perform a mediation analysis. First, we include the intent to churn dummy in full model (1) and estimate the model. We test for the relationship between intention to churn and actual churn. We find a significant positive coefficient (𝛽 = 2.433; 𝑝 < 0.001). We find that an increase of intention to churn from 0 to 1 a customer increases the likelihood of a customer churning by 11.23 times compared to a customer who did not indicate a high churn intention. This is a marginal increase of 41.59 percent in churn probability when a customer has a high intention of churning. We conclude that churn intention is a strong predictor of actual churn. As we hypothesize a mediation effect, we establish whether customer intent to churn acts as a mediating variable on actual customer churn through the various mediation paths. We follow the stepwise procedure to determine the total effect of the predictor variables on customer churn through the mediator (path c). Then, we determine the indirect effects of the predictor variables on through intended churn on actual customer churn (path a, path b).

Table 7

Mediation Analysis Significant Path Estimates

Variable Path Estimate P value Mediation Type

Age X → Y (c) -0.0495 <0.01 Partial Mediation X → Y (c’) -0.0308 <0.01 SX7B X → Y (c) -0.0251 <0.05 _Complete Mediation X → Y (c’) -0.0093 0.34 SX7C X → Y (c) 0.0556 <0.01 Partial Mediation X → Y (c’) 0.0452 <0.01

We find customer intentions to churn to act as a mediator for only three out of eighteen variables of the full model (1). We can only partially confirm H3a that the likelihood of customer churn is mediated by the intent of customers to churn.

4.5. Churn Intent Analysis

We now take churn intent as the dependent variable and look at the coefficients of path a of the mediation analysis (Appendix E). These estimates allow us to make inferences about the relationship between dissatisfaction of service quality and customer engagement variables on customer churn intent as hypothesized in H3.

Coefficient Signs. First, we look at the coefficient signs of economic determinants of satisfaction. We find a positive sign for aspect A, indicating a positive relationship between the dissatisfaction with the current health insurance price and intent to churn. We find negative coefficient signs for aspect B and G. This indicates a negative relationship between price increases and attractive competitive offerings and expressing churn intent. Aspect H also yields a positive sign, indicating a positive relationship between difficulties with paying and intention to churn.

Next, we look at the coefficient signs of service-related determinants of satisfaction. We find a positive sign for aspect D, indicating a negative relationship between dissatisfaction with the health insurer's current service and intention to churn.

Moreover, we find a significant positive sign for the engagement variable. This indicates that when a customer engages with the current health insurer the likelihood of expressing an intent to churn increases.

Lastly, we find a negative sign for aspect J, indicating a negative relationship between changes in personal health and the likelihood of expressing an intent to churn.

Marginal Effects. To interpret the model estimates, we now look at the marginal effects (table 8, appendix F). We only interpret significant results and relative to the reference category of ‘not important’.

For the economic determinants of satisfaction, we find that a three-unit increase in dissatisfaction with the insurer's price increases leads to an increase of churn intention by 8 percentage points. A one-unit increase in price dissatisfaction of the current health insurer decreases the intention to churn by 5.0, 4.7 and 7.67 percentage points. A one-unit increase in the importance of competitors' attractive offers decreases the intention to churn by 5.1 percentage points. A three unit increase in difficulties with paying decreased the intention to churn by 6.9 percentage points.

For service determinants of satisfaction, we find that a three-unit increase in

dissatisfaction with the health insurer's current service increases the intention to churn by 7.6 percentage points.

When a customer exhibits engagement behavior, the intention to churn increases by 5 percentage points. For customer characteristics, a unit increase in age decreases the intention to churn by 1.5, 6.2, and 6.0 percentage points. A unit increase in the level of education decreases the intention to churn by 8.2 percentage points.

Table 8

Marginal Effects Customer Intended Churn

Category Variable Marginal Effect SE P value Age Age 2 -0.0154 0.0189 0.41 Age 3 -0.0623 0.0195 <0.01 Age 4 -0.0601 0.0296 <0.05 Education Education 2 0.0222 0.0318 0.48 Education 3 0.0491 0.0286 0.08 Education 4 0.0816 0.0306 <0.01

Engagement Engagement Dummy 0.0505 0.0147 <0.01

Economic SX7A2 -0.0232 0.0218 0.28 SX7A3 -0.0101 0.0246 0.68 SX7A4 0.0807 0.0388 <0.05 SX7B 2 -0.0504 0.0180 <0.01 SX7B 3 -0.0477 0.0195 <0.01 SX7B 4 -0.0767 0.0179 <0.01 SX7G 2 -0.0512 0.0200 <0.01 SX7G 3 -0.0171 0.0225 0.44 SX7G 4 0.0128 0.0299 0.66 SX7H2 -0.0074 0.0200 0.70 SX7H3 -0.0222 0.0208 0.28 SX7H4 -0.0689 0.0180 <0.01 Service SX7D2 -0.0081 0.0240 0.73 SX7D3 0.0004 0.0262 0.98 SX7D4 0.0761 0.0323 <0.05 Other SX7I 2 0.0142 0.0247 0.56 SX7I 3 0.0303 0.0245 0.21 SX7I 4 0.1203 0.0421 <0.01 SX7J2 -0.0258 0.0196 0.18 SX7J3 -0.0098 0.0207 0.63 SX7J4 -0.0574 0.0193 <0.01

Categories: 2 = ‘slightly important’, 3 = ‘important’ 4 = ‘highly important’, age 2 = 36-55, age 3 = 56-75 age 4 = 76-113, income 2 <= €1700, income = 1700-3400, income 3 >3401, reference categories = ‘not important’, age 1 = ‘18-35’, income = ‘None’

4.6. Churn in Retrospect

switching or staying using the economic, service, and customer engagement related variables. The results are indicated in table 9.

Table 9

Switch or Stay in Retrospect

We find that when respondents look back at their decision-making process, economic determinants of satisfaction play a substantial role in customers' decision-making process. Looking at the relative importance of the churn determinants, we find that economic considerations account for 54% of the responses, while service-related considerations only account for 4,8% of the responses. Additionally, we find that customer engagement variables account for 3,8% of the responses. Non-engagement, such as visiting a user-friendly website, account for 2,1% of the responses. Lastly, a large number of respondents were indifferent to switching (16,5%) or stated no reason (1,1%) to stay or switch.

Overall, we find that pricing is the most important consideration for customers to stay or switch. In the customer churn logistic regression analysis, the most substantial effect on increased customer churn was found in a lower perceived price at a competitor. Thus, pricing being the overall most crucial reason for switching or staying is not surprising. Regarding

Churn Determinant Aspect Frequency

(Count) Percentage (%)

Economic Satisfaction

Price of the health insurance 241 23,7 Switching Discount Offered 94 9,2

Collective Discount 215 21,1

Service Satisfaction

Coverage of insurance 153 15,0

Terms of Service 41 4,0

Overall offered Service 8 0,8

Customer Engagement

Good Personal Advice 5 0,5

Advise by Financial Advisor 8 0,8 Advise by family & Friends

(WOM) 26 2,5

Customer

Non-Engagement User-Friendly Website 21 2,1

Other

Other financial products at

current insurer 27 2,7

Indifferent to Switching 168 16,5

Unstated other reason 11 1,1

service-related determinants, we find that coverage is the most important aspect of decision making. This is surprising, as the logistic regression analysis did not yield significant returns for the coverage predictor.

While the logistic regression analysis did not yield significant results for customer engagement variables, we do find that for 3,8% of the customers, this was their most important reason for staying or switching.

Lastly, we find that a total indifference to switching plays a substantial role, with 16.5% of respondents stating indifference as their most important decision-making factor.

5. Discussion

In this research, we built on the understanding of drivers of customer churn and customer churn intent in marketing science. We conducted research in the context of customer dissatisfaction in the Dutch health insurance market and examined the effect of economic and service-related determinants of satisfaction on customer churn. Moreover, we examined the effect of customer engagement and subsequent customer churn and customer churn intent. We included customer characteristics. Within our study, we set the following research question: “What are the drivers of customer switching behavior at the customer level?”

We subdivided the research question into satisfaction determinants, customer engagement, customer characteristics, and customer churn intent. This section serves to answer the research question and hypotheses using the results of section 4. The results are linked to the theoretical substance that was established in section 2. An overview of the hypotheses can be found in table 9.

5.1. Hypotheses