Customer churn analysis in the Dutch
Health Insurance market
2
Customer churn analysis in the Dutch
Health Insurance market
Assessing the influence of online and offline channel
usage.
by
Johan Lems
University of Groningen Faculty of Economics and Business Master thesis - MSc. Marketing Intelligence
June 26, 2017 Jadestraat 81-1 9743 HB Groningen 0610485429 j.a.lems@student.rug.nl S2196158
3
Management summary
In light of new legislations, insurance firms in the Dutch health insurance market face an increase in churning customers. As a result of increased market transparency, customers are the able to determine their alternatives better than before. On the other hand, firms are able to target customers more effectively. But since marketing resources are limited, firms are increasingly interested in how to allocate their resources across marketing activities in order to increase accountability of the marketing department. The impact of different marketing activities across different channels has been widely studied in marketing literature. By investigating different information channels in a unique setting, this study aimed to assess the influence of the use of different on- and offline information channels on the relation between the drivers of churn and customer churn behavior. In this paper, several binary logistic regression models are estimated to use a wide variety of explanatory variables to predict customer churn. By estimating different main effects and interaction effects models, the drivers of customer churn are examined and the influence of information channel usage and usage intensity is assessed. This study proves that financial benefits are one of the most important determinants of customer churn in the health insurance market. However, customer satisfaction, customer inertia and additional services are not proven to have a significant influence on churn behavior. While the results of this study show that the drivers of churn experience interaction effects. The influence of information channel usage is assessed and numerous interaction effects are identified. The proposed interaction effects proved their predictive ability in different logit prediction models in determining the customer churn behavior in the Dutch health insurance market.
Keywords: customer churn, health insurance market, behavioral loyalty, churn prediction
4
Acknowledgments
After a long period, the final hurdle of my academic experience in Groningen is here. During the past six years this master thesis has served as the ultimate goal that would mean the end of an amazing time in the city of Groningen.
Firstly, I would like to thank my thesis supervisor Edwin Kooge for guiding me through the process of writing my thesis. By providing me with helpful feedback he has been a great support. Secondly, I would like to thank many of my fellow students for countless interesting ideas, helpful insights and without whom my last year as a student at the University of Groningen would have been at lot less fun. In addition I would like to thank my friends and family for their unconditional support during the time I studied. Lastly, but most importantly, I would like to thank my girlfriend Lieke, who motivated and supported me throughout and who I admire for having an unbelievable amount of patience during the entire process.
Over the years I have been able to learn a lot, develop myself, met many new and interesting people and have made friends for a lifetime. But this thesis does not only mean the end of a great period of my life, but it also offers a lot of new opportunities for the future.
I hope that you will enjoy reading my master thesis.
Johan Lems
5
Table of Contents
Management summary ... 3 Table of Contents ... 5 1. Introduction ... 7 2. Theoretical framework ... 10 2.1 Customer churn ... 102.2 Drivers of customer churn ... 10
2.2.1 Financial benefits ... 11
2.2.2 Customer satisfaction ... 12
2.2.3 Customer inertia ... 12
2.2.4 Care package ... 13
2.3 Information channel usage ... 14
2.3.1 Online information channels ... 15
2.3.2 Offline information channels ... 16
2.4 Research framework ... 17 3. Methodology ... 18 3.1 Data description ... 18 3.2 Data preparation ... 20 3.2.1 Data aggregation ... 20 3.2.2 Weighting cases ... 21 3.2.3 Recoding variables ... 21 3.3 Descriptive statistics ... 23 3.4 Variables ... 23 3.4.1 Main effects ... 24 3.4.2 Interaction effects ... 25 3.4.3 Control variables ... 26 3.5 Research design ... 28 3.6 Research method ... 29 4. Model specification ... 31 5. Results ... 33 5.1 Main effects ... 33
5.1.1 Predicting household state ... 35
5.1.2 Predicting household churn ... 37
5.2 Interaction effects ... 38
5.2.1 Drivers of churn ... 38
5.2.2 Usage of on- and offline channels ... 38
5.2.3 Intensity of usage of on- and offline channels ... 39
6
6. Discussion ... 46
6.1 Conclusions ... 46
6.2 Limitations and future research ... 49
6.3 Managerial implications ... 51
References ... 52
Appendices ... 56
Appendix A: Variables ... 56
Appendix B: Descriptive statistics ... 58
Appendix C: Results – Interaction variables ... 59
7
1. Introduction
Over the last few years being able to measure the outcomes of marketing activities has become one of the most important subjects in marketing. Measuring and knowing the effect of marketing activities is vitally important in determining the accountability of the marketing department, which represents one of the main drivers of its influence within a company (Verhoef & Leeflang, 2009). Not only the accountability of the marketing department, but also the accountability of marketing activities has become an important subject. How to allocate marketing resources across channels and activities gets a lot of attention in the field of marketing research (Neslin & Shankar, 2009; Wiesel, Pauwels, & Arts, 2011).
The impact of different marketing activities and instruments across different channels has been widely studied in marketing literature. Where previous research focused on frequently purchased consumer goods (e.g., Deleersnyder, Geyskens, Gielens, & Dekimpe, 2002) or a business-to-business setting (e.g., Wiesel et al., 2011), this study focuses on the business-to-business-to-consumer services market. More specifically, this study investigates how customer churn influences the use of different information channels during the churn process in a unique setting, the Dutch health insurance market.
8
2015 (Vektis, 2017). Since customers in the Dutch health insurance market are not bound by multi-year contracts, they able to switch health insurer once a year facing low switching costs. As a result customers are able to churn more easily, especially this makes customers that do not choose to churn important for companies (Kumar & Shah, 2004). Consequently customers that do not churn, who can hence be seen as loyal customers, are recognized as a valuable asset in competitive markets (Srivastava, Shervani, & Fahey, 1998). Churning customers negatively affect companies in a number of ways. This not only leads to an immediate decrease in sales revenue, also, in case of services sold on a contractual basis, losing a customer is actually the end of the relationship with that customer (Risselada, Verhoef, & Bijmolt, 2010).
Not only the healthcare industry in the Netherlands has changed under the influence of legislation. With the Affordable Care Act (ACA) in the United States, the US government moved successfully uninsured people into coverage and improved retention rates in existing coverage programs by introducing more extended healthcare legislation (Graves & Nikpay, 2017). The ACA contributed to a decline in the uninsurance rate from 16.0 percent (48.6 million people) in 2010 to 9.1 percent (28.4 million) early 2016 (Graves & Nikpay, 2017). Besides, and more interesting for both health insurers and customers, the introduction of the ACA led to increased competition on the health insurance market in the United States and made customers within the healthcare industry more cost aware. This was mainly because of the increased transparency in the health insurance market (Jones & Greer, 2013).
9
churning process, marketing resources and marketing activities of Dutch health insurers, can be allocated more effectively, in order to decrease churning rates. Therefore the research question is:
How do drivers of customer churn predict churn in the Dutch health insurance market and how is this influenced by the usage of different on- and offline channels during the process?
This main question will be supported using the following sub questions: - What drives customer churn in the Dutch health insurance market? - Are their any interaction effects between the drivers of customer churn?
- What are the most important on- and offline information channels used in the churn
decision process?
- How does the use online information channels moderate the effect of the drivers of churn
on actual churn?
- How does the use offline information channels moderate the effect of the drivers of churn
on actual churn?
- How does the intensity of using on- and offline information channels moderate the effect
10
2. Theoretical framework
This chapter provides the study with the theoretical framework from which the hypotheses are derived. In order to do this the dependent variable, customer churn, is introduced and defined based on literature. Additionally, the concepts that influence the dependent variable are introduced. Based on a literature review, hypotheses that are tested in this study are formulated. Finally, the research framework is provided, which consists of a graphical representation of the hypothesized relationships between the different variables that are studied in this research.
2.1 Customer churn
Theories, definitions and concepts that involve customer churn, customer retention and customer loyalty are closely related to each other. Especially because customer loyalty has an important influence on attitudes towards customer churning behavior (Yang, 2014). In the specific situation of the Dutch health insurance market, as said before, customers only have the opportunity to churn once a year. This, e.g., in comparison to the Norwegian health insurance market, where customers are allowed to cancel their health insurance at any time of the year, not being limited to a certain moment or period (Günther et al., 2014). Therefore, customers who do not take the opportunity to churn can be seen as customers that stay loyal to the firm.
In current theories, customer loyalty is often described as consisting of two primary elements, namely attitudinal loyalty and behavioral loyalty (Chaudhuri & Hoibrook, 2001; Watson, Beck, Henderson, & Palmatier, 2015). This study focuses on behavioral loyalty. Attitudinal loyalty is the intention of a customer to stay loyal, whereas behavioral loyalty is the situation where a customer actually stays, e.g. by extending their contract. Measuring behavioral loyalty can be done by measuring the repeated purchase of product or service (Watson et al., 2015), which in this study can be measured by a customer that decided not to churn. Therefore customer churn in this study is defined as customers that churned from one health insurance firm to another during the fixed period that they were able to, and will be measured by either a ‘0’ (a customer that did not churn during the given period) or a ‘1’ (a customer that did churn during the given period).
2.2 Drivers of customer churn
11
customers that intent to churn, is first identifying the factors that drive a customer to churn (Tamaddoni Jahromi, Sepehri, Teimourpour, & Choobdar, 2010). In literature about customer loyalty, two different types of antecedents or drivers of churn can be distinguished; economic and social drivers (Bolton, Lemon, & Verhoef, 2004). In this chapter, different drivers of churn that can be classified as either economic or social drivers are explained. Economic drivers are explained as the economic attractiveness and financial benefits involved with churn. Social drivers are explained as the customer experience or attitude towards the health insurer, the extent to which the customer is satisfied, or the engagement of the customer in repetitive behavior explained by the inertia of the customer towards churning. In addition to economic and social drivers, the influence of additional services or supplementary healthcare packages is reviewed, as in the service industry offering additional services influences customer churn behavior (Wieringa & Verhoef, 2007).
2.2.1 Financial benefits
One of the most important and more obvious drivers of customer churning behavior is the financial benefit involved with churn. In service industries, economic attractiveness and financial benefits involved are identified as one of the critical drivers of the customer’s intention to churn (Yang, 2014). This is mainly because the customer’s evaluation of a service is in closely related with the economic costs of the service. Thus the evaluation of a service is closely related to the financial benefits that are involved with churn (Bolton & Lemon, 1999). From reviewing customer churn and loyalty studies in economic research, Yang (2014) concluded that churning in a service market is more likely to occur when the perceived benefits outweigh the perceived costs. The perceived benefits and cost consist of both economic and psychological aspects. An important determinant for churning intentions and churning behavior are the involved switching costs (Wieringa & Verhoef, 2007). Three different types of switching cost can be distinguished; procedural, financial and relational switching costs (Burnham, Frels, & Mahajan, 2003). As financial cost is an important part of switching cost, from an economic point of view, this suggest that the financial benefits is one of the drivers of customer churn behavior. The higher the financial benefits involved with churn, the more likely the churn behavior occurs. Thus the relation between financial benefits and customers is expected to be positive.
12
2.2.2 Customer satisfaction
The experiences of a customer with a product or service in the past can be a straightforward predictor of customer churn. Customers that experience higher levels of satisfaction with a product or service are more likely to have a higher usage of the product or service in the future (Bolton & Lemon, 1999). In a service environment, customer satisfaction drives the intentions of a customer to stay with or leave the service provider they were in the past. Therefore it is a powerful predictor of customer churning behavior (Burnham et al., 2003). In this study, based on the ideas of Verhoef (2003), customer satisfaction is defined as the extent to which the customer is satisfied with the service, expenses, fees, and coverage of their current health insurer. In addition, it also encompasses the extent to which the customer is satisfied with the interactions it has had with their current health insurer in the past. This suggests that customer satisfaction is one of the drivers of customer churning behavior, the more satisfied the customer is, the less likely it becomes that the customer is going to churn. Therefore the relation between customer satisfaction and customer churn is negative.
H2: Customer satisfaction is negatively related to customer churn.
2.2.3 Customer inertia
13
H3: Customer inertia is negatively related to customer churn.
2.2.4 Care package
All Dutch citizens are obliged to choose their own health insurance plan. This plan, in this research defined as the care package, consists of two different parts. The first part is the basic insurance, which is the same for every customer in the Dutch health insurance market. The second part consists of additional insurances or supplementary packages that can be added on top of the basic insurance. Because the basic insurance does not cover all health care, additional insurances or supplementary packages, like dental care, spectacles and contact lenses, physiotherapy or maternity care can be added to extend the customer health coverage (Ministry of Health, Welfare and Sport, 2017).
Since the basic insurance is the same for all the different Dutch health insurers, health insurances only become heterogeneous as a result of the additional insurances. The additional insurances and supplementary packages are, contrary to the basic insurance, not the same for every health insurer. Different insurers can offer different additional insurances or supplementary packages, which are priced at different levels and offer different benefits and coverage. In the service industry, offering additional services influences customer churn behavior (Wieringa & Verhoef, 2007). This means that health insurers can use the second part of the care package to distinguish themselves from their competitors.
This suggests that the care package is one of the drivers of customer churning behavior, the more additional insurances or supplementary packages are added to the care package, the more the differences between insurers increases as a result of more different packages and combinations of packages are offered. By offering additional services insurers tend to distinguish themselves from competitors. This makes the insurances offered by insurers more heterogenic. With increased heterogeneity, the more likely it becomes that the customer is going to churn. Since customers intend to look for the most attractive offer that applies to their personal preferences and situation Therefore the relation between the care package and customer churn is positive.
14
Based on the ideas of Bolton et al. (2004) and Wieringa & Verhoef (2007) four different drivers of churn are identified. The expected effects of these drivers on customer churn behavior are formulated in hypotheses 1, 2, 3 and 4. But in addition to the expected individual effects of the drivers, interaction effects among the different drivers might occur. Therefore it is expected that an interaction between two drivers might cause an increased effect of the drivers on customer churn behavior.
H5: The effect of one of the drivers on customer churn increases under the influence of another driver.
2.3 Information channel usage
During the process of constructing a churn prediction model for customers in the insurance market, it is not only important to predict whether a customer is likely to churn (Knott, Hayes, & Neslin, 2002), but also to predict how these customers are behaving in a multichannel shopping environment (Mau, Cvijikj, & Wagner, 2015). The implication of knowing how customers during the churning process behave makes it possible to anticipate on customer activities that predict churning behavior (Mau et al., 2015). Knowing when and where to target these customers increases the ability to allocate marketing resources more effectively across channels and activities (Neslin & Shankar, 2009; Wiesel et al., 2011). As the multichannel environment becomes a more common place where firms and customers interact, multichannel marketing activities become important tools to increase customer retention, increase customer value and decrease customer churn rates (Chang & Zhang, 2016).
15
distinction is made between on- and offline information channels when deteremining the influence of these channels on the drivers of churn. In addition a distinction is made between using an information channel (yes or no) and the intensity to which an information channels is used.
2.3.1 Online information channels
With the rise of online channels, customers are now empowered to gather information that is used in decision making processes more easy and quickly (Tamaddoni Jahromi et al., 2010). As a result, customers are able to make a more deliberate consideration with regard to different service providers and services. With increased transparency in the health insurance market (Jones & Greer, 2013), and the access to more information about additional services and alternatives influences customer churn behavior (Wieringa & Verhoef, 2007). While Chang & Zhang (2016) emphasize that the convenience of using online channels to gather information is of mere importance. Being able to access and gather information that supports customers during the churning behavior process is one of the most important functions of online information channels. An increased amount of information during the churning behavior process is expected to cause an increased effect of the drivers of churn on customer churning behavior. Therefore it is expected that the use and the usage intensity of online channels during the churning process will positively moderate the relation between the drivers of churn and customer churning behavior.
H6a: The usage of online channels increases the positive effect of financial benefits on customer churn.
H6b: The usage of online channels increases the negative effect of customer satisfaction on customer churn.
H6c: The usage of online channels increases the negative effect of customer inertia on customer churn.
H6d: The usage of online channels increases the positive effect of adding additional insurances or supplementary packages to the care package on customer churn.
16
H7b: The usage intensity of online channels increases the negative effect of customer satisfaction on customer churn.
H7c: The usage intensity of online channels increases the negative effect of customer inertia on customer churn.
H7d: The usage intensity of online channels increases the positive effect of adding additional insurances or supplementary packages to the care package on customer churn.
2.3.2 Offline information channels
In previous research, Chang & Zang (2016) explained that offline channels might be used to migrate customers from an inactive to an active state, where online channels are most commonly used to keep active customers in the active state. In their research, an inactive state is explained when a customer is not making use of the services or buying the products offered by the associated firm. While an active state is when a customer makes use of the services or buying products offered, or is considering to use it in de future (Chang & Zhang, 2016). Evidence shows that offline channels can help moving (potential) customers from an inactive to a more active state. Additionally, a more active state is expected to cause an increased effect of the drivers of churn on customer churning behavior. The use of offline channels during the churning process is therefore expected to positively moderate the relation between the drivers of churn and customer churning behavior. In addition it is also expected that the usage intensity of offline channels during the churning process positively moderates the relation between the drivers of churn an customer churning behavior.
H8a: The usage of offline channels increases the positive effect of financial benefits on customer churn.
H8b: The usage of offline channels increases the negative effect of customer satisfaction on customer churn.
H8c: The usage of offline channels increases the negative effect of customer inertia on customer churn.
17
H9a: The usage intensity of offline channels increases the positive effect of financial benefits on customer churn.
H9b: The usage intensity of offline channels increases the negative effect of customer satisfaction on customer churn.
H9c: The usage intensity of offline channels increases the negative effect of customer inertia on customer churn.
H9d: The usage intensity of offline channels increases the positive effect of adding additional insurances or supplementary packages to the care package on customer churn.
2.4 Research framework
The research framework for this study, including the hypotheses is shown in Figure 1.
18
3. Methodology
The dataset that is investigated in this study contains data on the Dutch health insurance market. The dataset was provided by a large marketing research institute specialized in consumer research. The dataset consists of survey data gathered from three different consumer panels on household level. Because no criteria information was offered in order to make a distinction between the different consumer panels in the dataset, differences between these panels will not be taken into consideration. The gathered data consists of information explaining the household situation of the consumers, the orientation period of the consumers including consumer search behavior, as well as purchase behavior. With regard to privacy issues, not all variables containing consumer specific information were available. An overview of all the variables that were available and used for this study is included in Appendix A.
This chapter provides the methodology that is used to prepare the data in order to test the hypotheses that are formed in the previous chapter. First, in order to achieve a good understanding of the dataset, the data will be described. Second, the variables in this study will be selected. In addition the data will be checked for missing values, outliers and partial responses. Lastly, the data will be prepared, by dealing with any inconsistencies in order to achieve better and more reliable results from the actual analysis.
3.1 Data description
19
The data that was gathered from the three consumer panels using monthly surveys is distributed across a total of five different datasets. These datasets contain information that is divided in five different topics; the structure of the data is shown in table 1.
Table 1. Data structure
Dataset Information
1. Household background information 2. Purchase behavior
3. Offline orientation 4. Search action 5. Online data
The used dataset is computed by taking relevant variables out of the different datasets and combine them into one analyzable dataset. The household identification number (HHID) is used to link and combine the variables to the right observations. The variables are selected based on their relevance with the conceptual framework and hypotheses, and can be categorized in the categories that are shown in table 2 below. A more extensive overview of the variables that are categorized within the different categories can be found in appendix A.
Table 2. Categories dataset
Category Information
1. General information observations 2. Household characteristics 3. Drivers of churn Financial benefits Customer satisfaction Customer inertia Care package
4. Information channel usage Online information channels Offline information channels
20
The variables in the category ‘Household characteristics’ contain information on household level as age of the housewife, household size, income, department size, the geographic district of the department and education level of the main earner of the household.
Finally, the variables in the categories ‘Drivers of churn’ and ‘Information channel usage’ contain numerous categorical and dummy variables. These variables are derived from questions from the surveys that contain information about the drivers of their churn behavior and information with regard to the information channels they used during the churning process.
3.2 Data preparation
3.2.1 Data aggregation
Currently the data is aggregated per observation. In order to be able to measure household behavior and being able to test the hypotheses in this study, the data will be aggregated on household level. The data will be aggregated using the unique household identification number as an identifier to restructure observations into variables. As a result, the dataset now contains 6,445 observations for 6,445 different households. After aggregating the data on household level, it can be observed what ‘type of buyer’ a certain household is. The possible options are either that the household ‘churned’ during the given period, they ‘actively extended’ with their current health insurer or they ‘passively extended’ with their current health insurer. An overview of the outcomes is shown below in table 3. When looking at the table, immediately two issues arise. First, the table shows that for one case, as results are measured with either a (1) ‘churned’, (2) ‘actively extended’ or (3) ‘passively extended’, two results are measured, as the total score is (5). In order to deal with this oddity, the case (with HHID 703993) is deleted from the dataset.
Table 3. Overview type buyers
Category Frequency % Valid
21
Second, a large amount of cases, namely 1,089 households, are observed as missing values. This may indicate that these observations are partial responses. Which means that these respondents did not answer all the questions in the survey, or dropped out in the process of filling in the survey. In most cases, partial responses do not immediately cause problems or concerns. But, since the question if the customer is a ‘churner’, ‘active extender’ or ‘passive extender’ is directly related to the dependent variable in this study, it concerns an essential question in order to get meaningful results. Therefore, these cases are deleted. As a consequence a total of 1,089 of the 6,444 cases are deleted. This results in 5,355 valid cases that can be analyzed.
3.2.2 Weighting cases
In order to check if the study is representable for the Dutch health insurance market, the distribution of the age of the housewife in households in the study is compared to the age distribution of the Dutch population. The results of this comparison are shown in table 4.
The results of this comparison show that the distribution in the dataset does not give a fair representation of the Dutch population. In order to achieve a fair representation and therefore a representable dataset, weight factors are used to weigh individual observations based on age categories. The distribution of the age categories is, and therefore the weight factors that are calculated, are based on population statistics of the Dutch Central Bureau for Statistics (CBS) for the year 2013 (CBS, 2017).
3.2.3 Recoding variables
The dataset that was structured for this study consists of numerous categorical, scale and binary variables. The variables have been recoded, in order to improve the usage of these variables and improve the interpretation of these variables in a latter stage. First, in order to match the weight
Table 4. Weight cases
Age category Dutch population % Dataset Desired distribution Weight factor 18-25 years 1,673,380 12.57 124 673 5,4270 26-35 years 2,021,472 15.18 608 813 1,3370 36-45 years 2,356,438 17.70 969 948 0,9779 46-55 years 2,505,387 18.81 1,008 1,008 0,8211 56-65 years 2,158,762 16.21 1,312 868 0,6617
Older than 65 years 2,600,643 19.53 1,046 1,046 0,9380
22
factors that are constructed, the scale variable age has been recoded into a categorical variable. The age variable has been recoded in six different categories, as shown in table 4 on the previous page. The socio-demographic variables that offer this study insight in the household characteristics of the observed households are also recoded. Variables like household income, education level, region, district size and household size are recoded. In the original dataset these variables consisted of 5 to 15 different categories, after recoding these variables contain between 3 to 6 categories, depending on the character of the variable.
Next, the independent variables in the drivers of churn category are addressed and recoded. The variables that are related to drivers of churn consist of variables that are either categorical or binary variables. Since the binary variables are already constructed in a way they can offer insights, these variables are not recoded. The categorical variables however contain contradictories in comparison to other variables in the dataset. Where most categorical variables in the dataset are constructed in a way that they are ordered from low to high, some are ordered from high to low. Therefore, the latter are recoded into a low to high order. After recoding, these variables now contain categories that are ordered from 0 ‘no influence’ to 3 ‘very high influence’, instead of 1 ‘very high influence’ to 4 ‘no influence’.
23
for inequalities in the number of surveys a household participated in. Now, a household that indicated the usage of an information channels once during three surveys has gotten an intensity of 0,33, while a household that indicated the usage of an information channels twice during two surveys has gotten an intensity of 1.
3.3 Descriptive statistics
After restructuring and aggregating the dataset, weighting cases and recoding a variety of variables, a comparison can be made between different groups within this study. The dataset consists of 5,355 observations on household level. The average age of the housewife of these households is 48 years old with a standard deviation of 16 years. On average, each household consists of 2.25 inhabitants and earns between 2,100 and 2,300 euros each month after taxes. The households are distributed across the Netherlands, 2,085 (38.9%) of the households come from the West of the Netherlands, 625 (11.7%) from the North, 1,252 (23.4%) from the East and 1,393 (26.0%) from the South of the Netherlands.
In this study two major distinctions can be made between groups. The first distinction is between households that were either actively or passively orienting during the churning process. The first group, consisting of 4,171 households, passively orientated. This means that they did not orientate, or hardly orientated to churn from their current health insurer to another. The second group, consisting of 1,184 households, has been actively orientating to change their health insurance during the observation period. In addition, within the group op households that has been actively orientating, a second distinction can be made between churners and non-churners. From the 1,184 households that had been actively orientating, and the total of 5,355 households in this study, 918 households switched to a new health insurer. Finally, 266 households that actively oriented did not switch, and extended their health insurance with their current health insurer. In this part of the study the descriptive statistics of these different groups have been briefly mentioned, a more extensive overview of the statistics per group can be found in appendix B.
3.4 Variables
24
3.4.1 Main effects
First the variables with regard to the main effects, the drivers of churn, are selected. The provided and constructed dataset offers a wide variety of variables that are derived from survey questions. Numerous variables in the dataset can be used to measure the effects of the drivers of churn identified in the literature review. In table 12 of appendix A an overview is given of the variables that might have measured the drivers of churn, the variables used for the analyses are selected from this overview. This overview is taken as a starting point for the variable selection. In order to specify a good model, the model needs to be simple to interpret and function as an extension of the conceptual framework to test the hypotheses and function as a prediction model. One of the criteria of a good model is that not too many variables should be selected, on one hand it has to be as easy as possible but on the other hand all necessary elements should be presented.
25
The same method is used to select the variable to measure customer satisfaction as a driver of churn. Two variables are identified as variables that can be related to customer satisfaction, performing a factor analysis is therefore considered inappropriate because it is not possible to delete a component in order to increase the performance of the factor. Thus, a binary regression is performed to test the individual effects of the variables. This test shows that only the variable CS1 has a significant effect on the dependent variable (𝛽 = -0.184, p = 0.000), and variable CS2 did not ((𝛽 = -0.08, p = 0.854). Therefore, it is decided to use FB3 to measure customer satisfaction.
Unfortunately, the variables that are identified in relation to customer inertia as driver of churn are measured within a biased sample. The surveys questions associated with customer inertia are only measured among the households that passively extended. As a result, it is not possible to include these variables in the analyses and prediction models, as this would give skewed results. This issue is addressed more extensively in the limitations of this study.
The binary variables, which are identified as variables related to care package, are also measured within a biased sample. Hence, the associated survey questions associated with care package are measured among the households that either switched or actively extended. Therefore, it is only possible to use the variable for the analyses that are performed on households in the active orientation state. This is explained more extensively in chapter 3.5, research design and 3.6, research method. Both CP1 and CP2 will be used as CP1 is a scale variable that includes the number of additional services included in the care package, and CP2 is a binary variable with the outcomes ‘only basic insurance’ and ‘basic and additional insurance’.
3.4.2 Interaction effects
26
variables that can interact with each other, the correlation test showed a lot of significant correlations. A total of 16 correlations is found. Since this does not give conclusive information on the different interaction effects, different models will be estimated and compared. Using the general-to-specific modeling approach a simple model without interaction effects and a full model with all possible interaction effects is estimated (Campos et al., 2005). Because it is expected that including all possible interaction effects in a single model is going to cause multi-collinearity issues, a third model will be estimated. This model is estimated with the use of the general-to-specific approach. With this approach, insignificant interaction variables are deleted until only the significant interactions are left. This way possible multi-collinearity issues are prevented. All three models are compared on predictive performance measures. With the general-to-specific approach a total of 14 interaction effects are identified, distributed over three different models. The results these effects are derived from can be found in chapter 5. In addition the interactions are included in an additional model specification, this can also be found in chapter 5. The method that is described is reciprocated to assess the moderating influence of the usage intensity of on- and offline channels. For the usage intensity variables, a full interaction model is estimated using the general-to-specific approach a second model is estimated with only significant interaction. The results of the possible interaction effects with regard to the usage intensity variables are included in chapter 5.2.
3.4.3 Control variables
The identified drivers and the usage of on- and offline channels are probably not the only factors that can have an effect on customer churning behavior. In order to accommodate possible other effects, control variables are included in the specified model. Based on previous research the following control variables are identified: Education and income, age, gender, household size and geography. An additional overview of the household characteristics used as control variables can be found in table 11 in appendix A.
Education and Income
27
education, customers that are higher educated are able to process new information more easily when making decisions. Because it is assumed that income is directly related to education level, customers with higher income levels are assumed to engage in information processing prior to the decision process (Homburg & Giering, 2001; Schaninger & Sciglimpaglia, 1981). Decision processes are essentially based on the evaluation of information. Hence with increased cognitive capabilities, customers with higher income and higher education levels are thus considered to be more likely to churn. For that reason, education and income are used as control variables.
Age
Previous research has shown that the desire of customers to churn decreases when the customer will increase in age (Wong, 2011). This can be a result of differences in information-processing abilities when evaluating different products and services, where these abilities decline with increasing age (Homburg & Giering, 2001). Thus, when predicting customer churn behavior, the age of the decision maker is used as a control variable.
Gender
Next, previous research found that compared to men, woman are more involved in purchasing activities (Homburg & Giering, 2001). Whereas a study by Melnyk, Van Osselaer, & Bijmolt (2009) states that differences between men and women are highly depended on the object that churning behavior is concerned with. Women occur to be more loyal to individuals, while men are likely to be more loyal to groups and companies. This is information suggests that women are more likely to churn, because they are less loyal to companies than men. Therefore, when predicting customer churn behavior, the gender of the decision maker is used as a control variable.
Household size
28
particular service occurs. With the increase of household size, small financial benefits can be multiplied and therefore become more significant. Thus, when predicting customer churn behavior, the household size is used as a control variable.
Geography
Finally, the influences of different geographic regions in the Netherlands will be taken into account. A possible reason for the influence of geographic regions might be that the access to health care is easier and greater in urban areas in comparison to rural areas (Pendzialek, Simic, & Stock, 2014). This is because different health insurances have contractual relations with different health care institutions and hospitals, in more rural areas the choice between health insurers is expected to be more limited as a result of more limited contractual relations. In addition, in line with customer inertia, it is expected that customers have long-term relationships with traditionally more regional health insurers. It is therefore expected that customers in the more rural areas (North, East and South) of the Netherlands are less likely to show churning behavior in comparison to customers in the more urban area (West) of the Netherlands. Therefore, when predicting customer churn behavior, the geographic region is used as a control variable.
3.5 Research design
The dependent variable in this study is measured on household level. Therefore, from this point forward in the specification and estimation part of the study, customers will be referred to as a household or households.
29
going to orientate itself, and after that the household decides to churn or not to churn 𝑌. Usually, in order to accommodate for these two stages, a hierarchical estimation method would be appropriate. The use of a nested binary regression, or nested logit, introduced by Foekens et al. (1997) would therefore be preferred. As a result of the fact that in this specific application, the second stage of the hierarchical model in the passive state only exists of a single choice, the usability of a nested logit is limited.
Hence, in order to estimate the probability of a customer is going to become active, and next the probability if a household is going to churn, binary logistic regression models, also known as logit models, will be used. By first calculating the utilities, the churn probabilities can be calculated. These in return, can be translated to a churn production. The first logit model will predict if either the household is going to become active (1) or is going to stay passive (0). The second logit model will predict whether the household is going to churn during the giving period (1) or the household is not going to churn during the given period (0). After estimation the model will provide insight in what variables are significantly influencing household and customer churn.
Since the goal of this study is to access what the influence of on- and offline information channel usage during the churn process is, the results of the estimated models will be descriptive. By explaining what the drivers of customer churn are and to what extend these drivers are moderated by on- and offline information channel usage.
3.6 Research method
The first logit model specified in this study considers the situation that a household chooses to either become active or stay passive. With the use of a binary logistic regression, in simple form, the binary choice model of household state can be explained by equation (1).
30
𝑆! = 1 if household 𝑖 decides to become active 0 if household 𝑖 decides to stay passive (1)
From Where 𝑃 𝑆𝑖 = 1 is the probability that a household decides to become active, and by definition 𝑃 𝑆𝑖= 0 is the probability that a household decides to stay passive. Because there are no other possible outcomes, equation (2) needs to be considered.
𝑃 𝑆𝑖 = 1 + 𝑃 𝑆𝑖= 0 = 1 (2)
The second logit model that is specified considers the situation that a household chooses to either churn or not to churn. With the use of a binary logistic regression (logit), in simple form, the binary choice model of household churn can be explained by equation (3).
𝑌! = 0 if household 𝑖 decides not to churn 1 if household 𝑖 decides to churn (3)
Where 𝑃 𝑌𝑖= 1 is the probability that a household decides to churn, and by definition
𝑃 𝑌𝑖 = 0 is the probability that a household decides not to churn. Like with equation (2), this situation also results in equation (4).
𝑃 𝑌𝑖 = 1 + 𝑃 𝑌𝑖 = 0 = 1 (4)
Based on the previous, with the structure of figure 2, whether a household is going to churn depends on the chosen state of the household. Therefore, the utility consists of two components. One component for the household state (𝑆) and one for the household churn decision (𝑌). From this point on forward, the utility that a household churns, given that the household has entered the active household stated, 𝑈!" is used to indicate the utility that a household churns.
Considering a linear model as proposed by Franses & Paap (2001) to estimate the utilities (𝑈!), the utility that a household decides to churn or not to churn, can be explained by equation (5).
31
Then the estimated utility that a household decides to churn or not to churn 𝑈!"! can be linked to the observed decision as shown in equation (6).
𝑌!=
1 if 𝑈!"! > 0,5
0 if 𝑈!"!≤ 0,5 (6)
Based on Verhoef et al. (2015) the logistic distribution function is chosen, from which the logit model will result and logistic regression is used to estimate the model. The specification of this model is preferred because it offers more mathematical convenience in comparison to e.g. a probit model (Verhoef et al., 2015). Because the logit model will include several independent variables, equation (7) can be generalized.
𝐹(𝑈!"!) = exp ( 𝑈!"!)
1 + exp( 𝑈!"! )
(7)
4. Model specification
In this chapter the models that are used to predict customer churn behavior in the Dutcher health insurance market is specified. This model is based on the hypotheses that are formulated in chapter 2 and the research method that is described in chapter 3. The models are specified as linear additive models, this is a model that is characterized by being linear in parameters and variables (Verhoef et al., 2015).
In equation (8) the first logit model is specified for the main effects of the drivers of churn on the probability that a household is going to enter an active state.
𝑈!! = 𝛼 + 𝛽!𝐹𝐵!+ 𝛽!𝐶𝑆! + 𝛽!𝐴𝐺𝐸! + 𝛽!𝐼𝑁𝐶! + 𝛽!𝐸𝐷𝑈! + 𝛽!𝐷𝐼𝑆! + 𝛽!𝐷𝐸𝑃!
+ 𝛽!𝐻𝐻𝑆! + 𝜀! (8)
32
including interaction effects is specified in chapter 5 after identifying significant interaction effects; in the full model only the significant interaction terms are included.
𝑈!|!! = 𝛼 + 𝛽!𝐹𝐵!+ 𝛽!𝐶𝑆! + 𝛽!𝐶𝑃! + 𝛽!𝐴𝐺𝐸!+ 𝛽!𝐼𝑁𝐶! + 𝛽!𝐸𝐷𝑈! + 𝛽!𝐷𝐼𝑆!
+ 𝛽!𝐷𝐸𝑃!+ 𝛽!𝐻𝐻𝑆! + 𝜀! (9)
𝑈!! The utility that household 𝑖 enters an active state
𝑈!|!! The utility that household 𝑖 is going to churn, given that it entered an active state
𝐹𝐵! Financial benefits variable for the prize at a different insurer is lower for household 𝑖
𝐶𝑆! Customer satisfaction variable for household 𝑖 to be unsatisfied with
the current insurer
𝐶𝑃! Care package variable for the number of additional care packages for household 𝑖
𝐴𝐺𝐸! Age category of the housewife for household 𝑖 𝐼𝑁𝐶! Income category for household 𝑖
𝐸𝐷𝑈! Education level for the main earner for household 𝑖 𝐷𝐼𝑆! Size of the municipality or district of household 𝑖 𝐷𝐸𝑃! Department region in the Netherlands of household 𝑖
33
5. Results
In this chapter the results of the estimated models and the different analyses are provided. First the results of the main effects are discussed. Next, the results of the relevant moderating effects are presented. Finally, the predictive validity is tested in order to show the relevance of the models used in the analysis.
5.1 Main effects
In this chapter the results of the estimated models and the different analyses are provided. First, the main effects of the drivers of churn on the probability that a household is going to enter an active state are discussed. The estimate parameter coefficients are shown in table 5 on the next page. Next, the main effects of the drivers of churn on the probability that a household is going to churn, given that it entered an active state are discussed. Table 6 on page 33 shows the estimated parameter coefficients of the simple model to predict household churn. In chapter 5.2 the results of the full model including the moderating effects are presented.
The results of the logistic regression that is performed to estimate the parameter coefficients are used to determine the utility and the probability of household behavior. For the results of the models the exponentiated estimated β’s are used to interpret the parameters more easily. The exponentiated β’s are the transformed β’s. With this transformation, the log odds ratios that are provided with β, are transformed in to the odd ratios. The odds ratio is shown with the use of equation (11), which is the probability an event that happens divided by the probability an event does not happen.
𝑂𝑑𝑑𝑠 𝑟𝑎𝑡𝑖𝑜 = 𝑒𝑥 𝑝 𝛽 = 𝑃 𝑆! 𝑜𝑟 𝑌! = 1 𝑃 𝑆! 𝑜𝑟 𝑌! = 0 =
𝑃!
1 − 𝑃! (11)
34
Table 5. Model 1 – Predicting household state
Logistic Regression P = .000
Category Variable Level Exp(β) S.E. Sig.
Drivers of Churn Financial benefits No role Reference category
Modest role 1.273 .135 .075*
Important role 1.786 .123 .000**
Very important role 2.454 .128 .000** Customer satisfaction No role Reference category
Modest role .921 .095 .39
Important role .777 .097 .001**
Very important role .629 .131 .000**
Customer Age 18-25 Reference category
characteristics 26-35 .411 .128 .000**
36-45 .342 .132 .000**
46-55 .372 .130 .000**
56-65 .343 .134 .000**
65* .264 .139 .000**
Income No income / not known Reference category
> 1500 1.019 .151 .900
1500 – 2500 .896 .146 .450
2500 – 3500 1.139 .149 .382
3500 < 1.194 .169 .293
Education No education Reference category
Special education 6,464 209.8 0.967 Primary education 13,342 209.8 0.964 Secondary education 14,607 209.8 0.964 Mid-level applied education 13,541 209.8 0.964 Higher professional education 16,813 209.8 0.963 Scientific education 15,222 209.8 0.963
District West Reference category
North .942 .121 .622
East .986 .093 .881
South .721 .096 .001**
Department size > 20.000 inhabitants Reference category
20.000 – 50.000 1.078 .128 .558
50.000 – 100.000 1.258 .135 .089*
100.000 < 1.186 .131 .192
Household size Small
Medium .994 .100 .954
Large 1.023 .167 .893
Constant 209.8 .960
35
5.1.1 Predicting household state
The first simple model, as presented in equation (8) in chapter 4, aims to predict the household state with regard to the churning process. A household can either enter an active or a passive state. The results of the estimation of this model show that the categorical variable that is used to represent the financial benefits associated with churn, where no role is set as reference category, has significant parameters. Financial benefits, as an important (p = .000) and very important (p = .000) role are significant on a significance level of 95% (p < .05). Modest role (p = .075) is significant on a significance level of 90% (p < .10). The probability increases with factor 2.454 (245.4%), if financial benefits play a very important role. The customer satisfaction variable, where no role is set as the reference category, also has significant parameters. Customer satisfaction as an important (p = .001) and very important (p = .000) role is significant. The exponentiated estimated β of the variable customer satisfaction is interpreted reversed, because the question where the variable customer satisfaction is derived from is asked in a negative way, ‘unsatisfied with the contracted hospitals/institutions associated with my current insurer’. Meaning, the probability increases with factor 1.371 (37.0%) if customer satisfaction plays a very important role.
36
Table 6. Model 2 – Predicting household churn – Simple model
Logistic Regression P = .000
Category Variable Level Exp(β) S.E. Sig.
Drivers of Churn Financial benefits No role Reference category
Modest role .794 .310 .458
Important role 2.442 .295 .002**
Very important role 2.879 .309 .001** Customer satisfaction No role Reference category
Modest role .686 .246 .126
Important role .680 .246 .116
Very important role .676 .336 .243
Care package .883 .047 .008**
Customer Age 18-25 Reference category
characteristics 26-35 .143 .518 .008**
36-45 .079 .511 .000**
46-55 .067 .503 .000**
56-65 .053 .502 .000**
65* .026 .508 .000**
Income No income / not known Reference category
> 1500 .456 .467 .094*
1500 – 2500 .412 .442 .049**
2500 – 3500 .575 .446 .214
3500 < .496 .479 .143
Education No education Reference category
Special education N.A. N.A. N.A.
Primary education 1.745 1.663 0.737 Secondary education .823 1.336 0.912 Mid-level applied education .807 1.334 0.872 Higher professional education .765 1.339 0.842 Scientific education .557 1.356 0.666
District West Reference category
North .702 .308 .250
East .892 .245 .640
South 1.010 .245 .967
Department size > 20.000 inhabitants Reference category
20.000 – 50.000 .647 .314 .166
50.000 – 100.000 .669 .336 .232
100.000 < 1.557 .341 .195
Household size Small Reference category
Medium .935 .257 .794
Large 1.202 .428 .667
Constant 124.32 1.49 .001**
37
5.1.2 Predicting household churn
The second simple model is a simplified version of the model presented in equation (10) in chapter 4. This model aims to predict the probability that a household 𝑖 is going to churn. The results of the estimation of this model show that the categorical variables that are used to represent financial benefits and care package as drivers of churn show significant estimates, as shown in table 6 on the previous page.
First, financial benefits, with no role set as the reference category, shows significant parameter estimates. Financial benefits as an important (p = .002) and very important (p = .001) role are significant. As a modest role shows no significance, this is contrary to the first simple model. The probability that a household will churn increases with factor 2.879 (287.9%) if financial benefits play a very important role. The care package variable also shows a significant parameter estimate (p = .008). Meaning that every additional package will cause a factor .883 (11.7%) decrease to the probability a household will churn. Customer satisfaction, with no role set as reference category, shows no significant parameter estimates. The probabilities for customer satisfaction as a modest (p = .126), important (p = .116) or very important role (p = .243) show no significant difference in comparison to the reference category.
The control variables education, district, department size and household size show no significant estimators in the simplified second model. However, the control variables age and income do show significant estimators. With the 18 – 25 age category set as the reference category, again all other categories show a significant decrease in probability when age increases (p = .000 for all categories except the 26 - 35 category, for 26 - 35 p = .008). A household in the category 56 – 65 years old has a decreased probability with the factor .053 (94.7%) to churn in comparison to the reference category. For the categorical income variable, with no income as reference category, showed for two categories significant estimators. The category < 1500 (p. = .094) on a 90% significance level, and the category 1500 – 2500 (p = .049) on a 95% significance level. Therefore, a household in the 1500 – 2500 category has a decreased probability with factor .575 (52.5%) to churn in comparison to the reference category.
38
5.2 Interaction effects
In this chapter the simple models that are used to estimate the main effects form the basis to find significant interactions effects. In order to assess the moderating effects of on- and offline information channels the significant interaction effects are used. The simple models are used to estimate three different forms of interaction effects. Interaction effects between the drivers of churn, between the usage of on- and offline information channels and the drivers of churn, and between the usage intensity of on- and offline information channels and the drivers of churn. 5.2.1 Drivers of churn
First the simple model that predicts the household state in equation (8), and the model that predicts household churn in equation (9) are estimated including interaction effects between the different drivers. Table 18 in appendix C shows a simplified output of the estimation where only the significant interaction effects are included. Interaction effects were only identified for the second simple model that predicts household churn. The output shows that a modest role for financial benefits interacts with a modest role for customer satisfaction (p = .090) at a 90% confidence level. In addition, a modest role for customer satisfaction interacts with care package (p = .043) at a 95% confidence level. Meaning that the effect of an additional care package on churn decreases with a factor .751 (24.9%) when customer satisfaction plays a modest role.
5.2.2 Usage of on- and offline channels
A full moderation model is estimated including all possible interaction effects between the drivers of churn and the usage of on- and offline information channels. Table 19 in appendix C shows the result of the significant interaction effects in a model where all insignificant interactions are deleted. The interactions that are found show that the usage of some channels moderates how the drivers of churn influence household churn. The significant interactions are therefore briefly presented below.
39
but this is a negative influence with factor .547. Meaning that reading magazines, papers or brochures decreases the influence of care package on the probability of churn with 45.3 In comparison to not reading magazines/papers/brochures. The interaction of the use of OFF4 (customer programs on radio or television) and financial benefits is also significant (p = .000), but has a negative influence with factor .019. This means that watching or listening to customer programs decreases the role of financial benefits as very important with 98.1% in comparison to not watching or listening to customer programs. On the other hand, the significant interaction (p = .017) between OFF6 (personal conversation with an advisor) and financial benefits as an important role is positive with factor 61.930. This means that a person conversation with an advisor increases the influence of financial benefits as an important role with 6093% in comparison to not having a personal conversation with an advisor. Finally, the interaction of OFF2 (reading offers) with customer satisfaction in a modest role is also significant (p = .052), but on a 90% confidence level. The interaction is positive with a factor 12.219, this means that reading offers increases the influence of customer satisfaction in a modest role on the probability of churn with 1122%.
5.2.3 Intensity of usage of on- and offline channels
To assess the moderating influence of the on- and offline information channels usage on the relation between the drivers of churn and household churn completely, the usage intensity interaction effects are estimated. Table 20 in appendix C shows the result of the significant interaction effects. The interactions found that the usage intensity of on- and offline information channels moderates how the drives of churn influence household churn.
The estimates show that the relation of financial benefits as a driver with household churn is moderated by 5 different usage intensities of information channels. The interaction of the intensity of ON2 (email contact with broker or insurer) and the important role of financial benefits is significant (p = .000) and has a negative influence and decreases the effect of financial benefits on household churn with 100% (factor .000). The interaction of intensity of OFF1 (reading magazines, papers or brochures) is significant and positive (p = .010, exp(β) = 7.281*103).
40
offers) are very contradicting, where the influence on financial benefits as an important role is significantly (p = .065 on a 90% confidence level) positive, the influence on financial benefits as a very important role and OFF2 is significantly (p = .020) negative. The interaction of the intensity of OFF3 (discussing with friends and family) is significant (p = .029) and positive (exp(β) = 171.4) with the modest role of financial benefits. This means that when the usage intensity increases with 1, the effect of the modest role of financial benefits will increase with 17040%. Lastly, the interaction with the intensity of OFF4 (customer programs on radio or television) is significant (p = .000) and negative with factor .000. This means that an increase of usage intensity with 1, the effect of financial benefits on household churn decreases with 100%. In addition, customer satisfaction significantly interacts with two usage intensities. The usage intensity of the usage of website (ON1) and e-mail contact with a broker or insurer (ON2). The intensity of ON1 has a significant (p = .035) negative interaction with factor .000. This means that an increase in usage intensity for ON1 with 1 will decrease the effect of customer satisfaction on churn with 100%. The intensity of ON2 has a significant positive interaction with extremely high factors (exp(β) respectively 2.578 *107, 1.958 *1010 and 1.595 *105 for modest, important
and the very important role).
Now that possible interaction effects are identified, a full significant interaction model can be specified. An overview of the interactions effects is given in table 7 below.
Table 7. Identified interaction effects
(1) (2) (3)
Simple model Full model with usage interaction effects
