THE IMPORTANCE OF PRICE IN
MODELLING AND PREDICTING CHURN IN
THE DUTCH HEALTH CARE INDUSTRY
2
THE IMPORTANCE OF PRICE
IN MODELLING
AND PREDICTING CHURN
IN THE DUTCH HEALTH CARE INDUSTRY
Master Thesis by Mark Pijper
University of Groningen Faculty of Economics and Business
MSc Marketing July 2017 Vrydemalaan 654 9713 WZ Groningen (06) 55526259 m.e.pijper@student.rug.nl Student number: 3031349 Supervisor: N. (Natasha) Walk
3
Preface
By deciding to start with the Master Marketing Intelligence I have found out that this direction suits me. Working on analytical challenges has certainly sparked my enthusiasm. Before I started the Master I was not quite sure what kind of job I would imagine myself doing in the future, and during the whole process I discovered my passion for data analytics. At the start of the Master I had no knowledge of how to work in R and now I can confidently say I’m quite proficient in it. This can be contributed to the challenging datasets I had to work with, and also to my fellow students with whom I discovered R.
I want to thank my supervisor Natasha Walk for her valuable insights and feedback during the process, and my colleagues from my thesis group for their input. I would also like to thank my friends from the Master for keeping me motivated by dragging me to the campus with them, and for the help they have provided me with when necessary. In particular I want to express my gratitude to Shu Han Chuang, Laura Mensink, Diana Solfanelli, & Wisse Smit for taking the time to read my thesis and provide detailed feedback at the final moment.
Honestly I cannot say I enjoy writing reports like this, but I did thoroughly enjoy the analytical challenge it provided and I am proud of the final product. Goodbye Master Marketing, I enjoyed my time here.
Mark Pijper,
4
Management summary
Competition between firms remains fierce with markets becoming more saturated, and the effort for gaining customers is becoming more intense, resulting in a situation in which firms are forced to spend more resources on customer retention (Roos & Gustafsson, 2007). Therefore, retaining customers is becoming more important for managers, which emphasizes the need for more efficient methods to predict customer churn. A small improvement in customer retention would allow the company to generate more long-term revenue. Various studies have identified the causes of customer churn, of most of which were conducted in the mobile service industry. These studies showed that the causes are related to economic value (Lim et al., 2006), customer satisfaction/dissatisfaction (Ahn et al., 2006), demographic factors (Ferreira et al., 2004), and the quality of service (Mozer et al., 2000. Economic value refers to the pricing element. The quality of service is a part of customer satisfaction/dissatisfaction, as a perception of high quality of service will lead to satisfied customers (Jajaee & Ahmad, 2012). Which of these causes is the most powerful driver of churn? The purpose of this paper is to identify whether price is the most powerful driver of churn and build a model which is able to accurately predict churn in the health insurance sector. An increase in the price of the premium has been sighted over the years in this sector, and the percentage of churning customers has also been increasing (Vektis, 2016), motivating the choice for price as the expected largest driver of churn. Identifying the most powerful driver of churn will allow managers in this sector to allocate their resources in a more effective way. Customer churn analysis literature in the health insurance market is scarce, and this study will contribute to that gap. A dataset containing switching behavior shown by households concerning their health insurance has been used to conduct the different analyses which are described in the methodology.
5 performed along with causal mediation analysis by nonparametric bootstrapping to estimate the different models, and identify the biggest driver of customer churn.
This study found that perceived price is the biggest driver in predicting churn, directly and through churn intention. The other main predictor, customer dissatisfaction, was not significant and could therefore not be compared to perceived price. The model accommodating the factor for different health insurers is the strongest model, but due to the nature of the data the sample size used for this model is smaller than for the other models. Therefore, the model is not suited for churn prediction in the current sample size. Adding the aspect of the collective scheme to the base model as a moderating effect results in the strongest model for churn prediction purposes, as opposed to adding all the variables of the conceptual model into a complete model. The model accommodating the collective scheme uses the complete sample size, and is therefore better suited for churn prediction than the model accommodating the factor for the different health insurers.
The effect of perceived price on churn behavior differs for smaller insurers and the insurer ‘Achmea’, when using insurer ‘CZ’ as the baseline. Both ‘Achmea’ and ‘CZ’ are part of the five largest insurers in the Netherlands. This effect is not mediated through churn intention. A change in the dissatisfaction of the customer negatively influences the effect of perceived price on churn, fully mediated through churn intention. This also holds true for the presence of a collective scheme, a change in the personal situation of a customer and a change in the health of a customer. The moderating effect of customer dissatisfaction on perceived price is negative, which was not expected. Different machine learning algorithms have been developed for the strongest model. The algorithm for the logistic regression is the best overall classifier by taking the simplicity of the technique and the performance into account.
6
Table of contents
1. Introduction ... 8
2. Theoretical framework ... 11
2.1 Dutch Health insurance sector ... 11
2.2 Conceptual model ... 13
2.3 Customer churn ... 14
2.4 Independent variable ... 15
2.5 Covariates ... 16
2.6 Moderators ... 17
2.7 Mediation effects of churn intention ... 19
2.8 Churn model approaches ... 20
3. Research design ... 21
3.1 Description ... 21
3.2 Measurement of variables ... 22
3.3 Models ... 26
3.4 Steps to reject the null-hypotheses and build the model with the best performance ... 27
3.5 Prediction of churn ... 28
4. Results ... 31
4.1 Correlation table ... 31
4.2 Logistic regression ... 32
4.3 Mediation analysis ... 34
4.4 Comparing the models and determining the best model ... 35
4.5 Machine learning ... 38
5. Conclusion ... 40
References ... 47
Appendix A – Factor analysis ... 54
Appendix B – Equations for the different hypotheses ... 58
Appendix C – Mediation analysis ... 60
Appendix D – Churn intention as dependent variable ... 64
7
ABSTRACT:
The retention of customers is becoming more important for managers, which emphasizes the need for more efficient methods to predict customer churn. The purpose of this paper is to identify whether price is the most powerful driver of churn and build a model which is able to accurately predict churn in the health insurance sector. This study examined churn prediction by using the perception towards price, customer dissatisfaction and demographic variables as predictors. Furthermore, the different insurers, the usage of a collective scheme, a change in health, a change in personal situation of the customer, and customer dissatisfaction have been added as moderators. The model has been made more complex and stronger by adding churn intention as a mediating variable. Logistic regressions have been performed along with causal mediation analysis by nonparametric bootstrapping to estimate the different models and identify the biggest driver of customer churn. This study found that price is the biggest driver in predicting churn both directly and through churn intention. Different machine learning algorithms have been developed for the strongest model, with the logit model being the best overall classifier by taking the simplicity of the technique and the performance into account. The findings imply an expansion of customer churn analysis and churn prediction by using a method which allows for the estimation of a logit model, while enabling the estimation of the indirect mediating effects. The resulting model can be used by health insurers to predict churn and therefore advance competitive advantage. This study contributes to current research by accommodating both mediating and moderating effects in the model, and implementing the specific method in customer churn analysis.
Keywords: Churn prediction, machine learning, health insurance, logit model,
8
1. Introduction
Customer churn is becoming a significant problem for firms in different sectors, and is developing into one of the prime challenges firms have to face worldwide (Chandar, Laha, & Krishna, 2006). It is becoming an industry-wide belief that the best core marketing strategy for the future is to retain existing customers and avoid customer churn (Kim, Park, & Jeong, 2004; Kim & Yoon, 2004). Retaining clients costs five to six times less than attracting new customers (Athanassopoulos, 2000; Bhattacharya, 1998; Colgate & Danaher, 2000; Rasmusson, 1999). Identifying and predicting which customers are likely to churn will allow firms to manage churn, as the firms are then able to target incentives to these customers, resulting in higher customer retention and lower customer churn. It will save firms a large amount of money if they are able to more effectively target customers who are at risk to churn, because the firm knows which customers to focus its efforts on. Competition between firms remains fierce with markets becoming more saturated, and the effort for gaining customers is becoming more intense, resulting in a situation in which firms are forced to spend more resources on customer retention (Roos & Gustafsson, 2007). According to Bijmolt et al. (2010) the prevention of customer churn is the main focus of customer retention. A small change in the retention rate can result in significant impact on businesses (Van den Poel & Larivière, 2004). The research of Van den Poel & Larivière (2004) used a real-life example of actual financial services companies to illustrate that an increase in retention rate of just one percentage point may result in substantial profit increases. Considering a time period of 25 years and a 6 percent discount rate in this example, the improvement in 1 percent of retention would increase the contribution by 27.5 million Euros. This can be seen as a dramatic impact, because a small improvement in customer retention would allow the company to generate more long-term revenue. Concluding, retaining customers has become a critical issue for managers (Buckinx & Van der Poel, 2005), resulting in companies being eager to build more efficient methods to predict customer behavior.
9 service industry. Economic value refers to the pricing element. The quality of service is a part of customer satisfaction/dissatisfaction, as a perception of high quality of service will lead to satisfied customers (Jajaee & Ahmad, 2012).
Customer churn is also an issue in the healthcare industry: the amount of people that switched healthcare insurer has increased in 2013 (Reitsma-van Rooijen, 2013). This trend has also occurred in the years up to 2013 (Reitsma-van Rooijen et al., 2012; Reitsma-van Rooijen et al., 2011; Vos en De Jong, 2009; De Jong, 2008; De Jong en Groenewegen, 2007). Churn behavior is increasing from 3.5 percent in 2007 to 5.5 percent in 2016 (Vektis, 2016). The most important reason that people switch healthcare insurer in the Netherlands is because of the level of the total premium (Reitsma-van Rooijen, 2013; Maarse et al., 2016). Dutch customers are allowed to switch health insurer once a year, thus encouraging competition (Maarse et al., 2016). All applicants must be accepted by insurers for the basic scheme, and apply community-rating.
The expectation is that price of the premium is the biggest driver for customer churn (Reitsma-van Rooijen, 2013), and therefore the biggest predictor for churn. It is unclear whether this is the same across all health insurers. If the price of the premium is indeed the biggest predictor of customer churn, then it would be possible to create a general churn prediction model that can be applied to all health insurers based on price. An alternative could be that there are different types of customers subscribed to specific health insurers, resulting in different results and therefore the need for different models. A goal of this research is to find out whether the specific insurer matters in building prediction models. This research also takes into account the presence of a collective scheme for, as well as a shift in the personal situation or the health situation of a customer as these are also expected to have an influence on customer churn.
10 specific method in customer churn analysis. The following research questions describe what questions will be answered in this study:
What effect does perceived price have on churn behavior in the health insurance sector
in the Netherlands?
Does a change in the dissatisfaction of the customer influence this effect? Is this effect equal for all health insurers, or are there differences?
Is this effect equal for customers who are part of a group insurance, or are there
differences?
Does a change in the personal situation of a customer influence this effect? Does a sudden change in the health of a customer influence this effect?
What is the most suitable machine learning approach to predict churn in the health
sector?
Neslin et al. (2006) indicate that multiple steps in the churn prediction process have an impact on its success, but they strongly suggest putting the focus on the prediction technique. This is due to its huge impact on the return on investment of subsequent marketing actions.
11
2. Theoretical framework
In this chapter the conceptual framework is presented for the influence of price perception on customer churn. Information concerning the Dutch Health insurance sector is provided, along with the theoretical framework describing all the constructs which are shown in the
conceptual model.
2.1 Dutch Health insurance sector
In the United States, the expected percent change in marketing budgets from February 2017 to 2018 is expected to be 10.1 percent higher in the health care industry (The CMO Survey, February 2017). The marketing spending as percent of company revenues is 10.2 percent for this sector. The expectation is that this trend will also be sighted in the Netherlands. Still, different countries have different healthcare policies, although the empirical results of this research should also be of interest to readers outside of the Netherlands. In order to fully comprehend the concepts of this research, an introduction to the concept of the Dutch Health insurance sector has been given below.
The Dutch healthcare system is different to other countries, as a major reform of the health care system happened in 2006, and according to the Health Consumer Powerhouse the Netherlands is using a so-called ‘chaos system’ (Euro Health Consumer Index, 2015). The ‘chaos system’ allows patients a great degree of freedom concerning where they want to subscribe for their health insurance and where they get it. This means that the new ‘chaos system’ will allow customers to switch from one insurer to another more easily, which should drive insurers to increase the efficiency of the health care provision in order to retain their customers (Boonen, Laske-Aldershof, & Schut, 2016). Thus, the result of this new system is more competition among health insurers, as it is unsure whether customers will be retained (Günther et al., 2014).
12 private health insurers, and in some cases the employees’ premiums are paid as an employment benefit. Employees can partake in this ‘collective insurance’. An important feature of this healthcare system is that premiums may not be related to age or health status. Table 1 describes the average price of the premium, which has been calculated by using the average of all 66 policies over the previous four years. The percentage increase compared to the previous year is shown after the price. Own risk is the amount that the insured have to pay for themselves when they need medical care that is covered by the basic insurance. This amount can also change yearly. Aside from the obligated own risk, customers can also choose freely for extra own risk, and the higher the freely chosen own risk is, the less premium the customer pays. As the costs are becoming higher for customers, the focus on price is rising, which is why it is expected that pricing is the most important driver of churn. The Minister of Health is responsible for managing health care expenditures. Although regulated computation is assumed to be an effective instrument for cost control, a yearly global budget (the macro-budget) is imposed to cap expenditures (Maarse et al., 2016). This means competition is combined with global budgeting. The insurers’ premium rate varies largely, as in 2014 the difference between the highest-priced and lowest-priced nominal premium was 30 percent (BS Health Consultancy, 2014; Reitsma-Van Rooijen & De Jong, 2014). A drop in customer mobility was sighted in 2014, which is most likely due to a decrease in the average nominal premium (Vektis, 2014).
Average premium Own risk
2014 €93.89 €360
2015 €98.96 (+5.4%) €375 (+4.2%)
2016 €101.67 (+2.7%) €385 (+2.7%)
2017 €109.62 (+7.8%) €385 (+0%)
Table 1 – Average premium and own risk
13
2.2 Conceptual model
The conceptual model is presented in figure 1. This study will use perceived price to predict churn. Customer dissatisfaction has been added to the model as a covariate to predict churn. Demographic factors are also added as covariates in order to allow for more accurate predictions.
Different health insurers are also taken into account as moderator, as they offer different insurances for customers and also have different sizes. Furthermore, as described earlier, a large amount of customers is part of a collective scheme, which needs to be part of the model. Finally, a customer might experience a change in their personal situation, or in their health, which both might weaken the effect of the price. Customer dissatisfaction has also been added to the model as a moderator, because an interaction effect is expected. More satisfied customers will behave differently than dissatisfied customers concerning price. The intention to churn is a mediating variable in this model between the variables that influence customer churn and the actual churning.
All the different variables will be described in the next paragraphs. As displayed in the conceptual model, all main hypotheses have mediating counterparts which describe the same effect, but through the mediating variable.
Independent variable:
Negative perceived price of the premium Mediator: Churn intention Dependent variable: Customer churn
+
+
+
Moderators: A. Customer dissatisfaction B. Health insurer C. Collective scheme D. Change in health consumer E. Change in personal situation consumerCovariates A. Customer dissatisfaction B. Demographics customer H1' H1 H2 H2'
14
2.3 Customer churn
Kamakura et al. (2005) states that “churn refers to the tendency for customers to defect or crease business with a company”. Customer churn modeling can be seen as a binary classification problem, as the customer either churned or stayed with the company. There are two types of targeted approaches to managing customer churn: reactive and proactive (Burez and Van den Poel, 2007). In the case that a company adopts a reactive approach, the company waits until the customer will ask the company to cancel the service relationship, enacting the company to offer incentives to stay. In the case that a company adopts a proactive approach, it will try to identify customers who are likely to churn before they do so. In this situation the customer will provide special incentives to keep the customer from churning. Proactive programs have lower costs, but the churn predictions need to be accurate, as incentive money will be spent on the wrong customers in case of inaccuracy. For this reason, building a customer-churn prediction model which is as accurately as possible is key (Burez & Van den Poel, 2007; Van den Poel & Larivière, 2004), which is a goal of this study. Numerous predictive-modeling techniques are available for this, and the selection of customers most prone to churn can be done with these data-mining techniques (Hung et al., 2005). Multiple predictive-modeling techniques will be described at the end of this chapter.
According to Ahn et al. (2006), customer churn behavior is a big issue in the services industry, and Neslin et al. (2006) elaborates on this by stating that the health care industry is one of the industries that face significant problems with churning customers. This is in line with the developments in the Dutch health care industries which has been described earlier.
15 customers with choosing a plan that suits them better. However, Ascarza, Iyengar, & Schleicher (2015) did research involving customers of a wireless telecommunication service, and the results indicated that being proactive and encouraging customers to switch to better tariffs can increase customer churn. Of the control group, who received no encouragement, 6 percent churned. Of the treatment group 10 percent churned. An explanation provided by the study for this finding was a change in customer inertia, because of the intervention by the firm. Another explanation was that the encouragement to switch plans also emphasizes past usage. Both explanations were supported by the data. An important side note is that in the research over-consuming customers were encouraged to upgrade to a better plan. It is not known what would happen when it would be best for customers to downgrade their service contracts. However, this does raise the question: How can churn be prevented if giving recommendations in terms of price to customers has the opposite effect? There are more factors that contribute to churning, as was described before. Before focusing on effectively preventing churn by means of a retention program, the first matter that needs to be addressed is determining whether price is the largest driver in predicting churn or not.
2.4 Independent variable Perceived price of the premium
16 As stated before, the most important reason for switching healthcare insurer in the Netherlands is because of the level of the total premium (Reitsma-van Rooijen, 2013; Maarse et al., 2016), a finding that has not been analyzed in a churn prediction setting before. Since costs are becoming higher for customers, their focus on price is rising and therefore, it is expected that pricing is the most important driver of churn.
H1: Negative perceived price of the premium has a positive influence on customer churn
2.5 Covariates
2.5.1 Customer dissatisfaction
Customer satisfaction can be seen as customer reaction to the state of fulfilment, and customer judgment of the fulfilled state (Oliver, 1997). Fornell (1992) states that customer satisfaction heightens customer loyalty and prevents customer churn, which in turn lowers customers’ price sensitivity, reduces the costs of failed marketing and of new customer creation, reduces operation costs due to the increase in number of customers, improves the effectiveness of advertising, and enhances business reputation. The customers’ own perceptions of service quality are the main factor in determining customer satisfaction (Zeithamal & Bitner, 1996). Service quality, according to Parasuraman, Zeithamal, & Berry (1998), can be defined as the customers’ satisfaction or dissatisfaction which has been formed by their experience of purchase and use of the service. Ahn et al. (2006) conducted a multinomial logistic regression on subscriber data in Korea for the telecom industry, and discovered that customer dissatisfaction affects churn. This study focuses on price, thus customer dissatisfaction has been included as a covariate to predict churn and to compare the size of the effect with price perception.
2.5.2 Demographics customer
17 parents chose different health plans for their newly enrolled children, showing the relevance of adding income and household size.
2.6 Moderators
2.6.1. Customer dissatisfaction
Anderson (1996) has found a positive association between changes in customer satisfaction and changes in price tolerance, but a negative association between the level of customer satisfaction and the level of price tolerance. From the perspective of firms, this implies that higher customer dissatisfaction is likely to increase price sensitivity. To account for the relationship between these two constructs, customer dissatisfaction will also be used as an interaction effect to predict the effect of price on churn.
H2a: The effect of negative perceived price of the premium on customer churn is stronger if the level of customer dissatisfaction is higher.
2.6.2. Health insurer
The Dutch health care system consists of nine different independent concerns which harbor one or multiple different health insurers or proxies. In 2016 the four biggest concerns, which are Achmea, CZ, VGZ and Menzis, have 88 percent of the people that are insured as customer (NZa, 2016). The five other concerns have gotten more market share over the years, but are still way behind the top four. Customers can choose from around 50 health insurers. A customer has churned when it has switched from one health insurer to the other, even if the switch was made to a health insurer of the same independent concern.
In 2006 healthcare policy has taken a major step in the gradual transition from supply-side government regulation towards regulated competition (Loozen, 2015).
It was described before that the contents of the basic insurance is decided by the Dutch government, and is equal for everyone. Still every health insurer will calculate a different premium, which is done because every health insurer will make different (pricing)agreements with health providers. Different health insurers also have different financial reserves, which are also taken into account when deciding on the price of the premium.
18 tend to equate better quality with lower costs, measure quality solely in financial terms and show little to no interest in better quality with higher costs. This shows that providers and insurers might have different conceptions of quality of care and quality-cost balance. Some insurers also strive to concentrate high-complexity services in only a limited number of hospitals, but these initiatives are meeting resistance (Maarse et al., 2016).
Anderson (1996) indicates that greater competitiveness in an industry is associated with lower price tolerance, therefore greater price sensitivity.
H2b: The effect of negative perceived price of the premium on customer churn differs across different health insurers.
2.6.3. Collective scheme
Health insurance legislation allows for collective schemes with a maximum premium discount of 10 percent (Maarse, 2016). Collective plans need to be taken into consideration, as a high percentage of customers are in a collective scheme (69 percent in 2014), meaning the plans play a prominent and competitive role in this industry (Vektis, 2014). The most renowned collectivity schemes are the employee collectivities. These are launched by an employer and are meant for their employees (NZa, 2016).
According to Maarse et al. (2016) there is controversy about the fairness of collective schemes, as it is argued that discounts obtained through collective bargaining are compensated by higher premiums being paid by persons who do not either qualify or opt for a collective scheme. The relationship between individual and collective schemes has been described as a zero-sum game (KPMG/Plexus, 2014), thus undermining solidarity. Another criticism is that collective schemes compromise transparency in health insurance. Several studies have found that prices matter for retirees, but substantially less than for active employees (Buchmueller et al, 2013). Finally, it has been shown that having group insurances is likely to have a negative impact on customers’ willingness to churn, as these contracts are often concluded for multiple years (van der Maat & de Jong, 2009).
19
2.6.4 Change in health of customer
In the case that a customer perceives a positive or negative change in their health, it can be assumed that this might lead to the search for a health insurance package that would fit them better, resulting in churn. For example, a customer could have injured his/her back, resulting in a high amount of physical therapy visits, where this was not needed beforehand. The model will account for this phenomenon. The effect of perceived price on churn is expected to weaken if a change of health is perceived, as price would play a smaller role in this situation. H2d: The effect of negative perceived price of the premium on customer churn is weaker if a change in the health of a customer is perceived.
2.6.5 Change in personal situation of customer
It might also be the case that the personal situation of the customer changes positively or negatively. Factors like a change in income or a change in the size of the family are assumed to have an influence on churning behavior. The demographics that are not dynamic in nature (i.e. income and household size) are accounted by including a variable which describes a change in the personal situation in the model. In 2013, 10 percent of the customers churned because of a change in their family situation (Reitsma-van Rooijen, 2013). The model will need to account for this change in order to determine the importance of price. The effect of perceived price on churn is expected to weaken if a change in the personal situation is perceived, as the role of price would be smaller in this situation.
H2e: The effect of negative perceived price of the premium on customer churn is weaker if a change in the personal situation of a customer is perceived.
2.7 Mediation effects of churn intention
20 having an influence on customer churn and the actual occurrence of customer churn will be mediated by churn intention. Madden, Savage & Coble-Neal (1999) have done research in the Internet service provider market, and found that respondents indicated that affordability of the subscription is an important factor in their intention to churn. The expectation is that this finding could also be extended to the health insurance market. Prior research in the health insurance market has included customer churn in their models (Günther et al., 2014), but not the intention to churn, which is why an assumption has to be made. Therefore, it is assumed that the effect of perceived price on customer churn is mediated by churn intention. All the moderating effects that have been described in chapter 2.5 will also be included in the indirect effect of perceived price on customer churn, through churn intention. It has been stated that all these moderators have an effect on perceived price, however, it has to be researched whether these effects are mediated through churn intention.
H1’: The effect of negative perceived price of the premium on customer churn is mediated by churn intention.
H2a’: The effect of negative perceived price of the premium on customer churn when the customer is dissatisfied is mediated by churn intention.
H2b’: The effect of negative perceived price of the premium on customer churn across different health insurers is mediated by churn intention.
H2c’: The effect of negative perceived price of the premium on customer churn when a collective scheme is involved is mediated by churn intention.
H2d’: The effect of negative perceived price of the premium on customer churn when a change in the health of the customer is perceived is mediated by churn intention.
H2e’: The effect of negative perceived price of the premium on customer churn when a change in the personal situation of the customer is perceived is mediated by churn intention.
2.8 Churn model approaches
21 examining different machine learning methods. In the research of Vafeiadis et al. (2015) the boosted Support Vector Machine was the overall best classifier with an accuracy of almost 97 percent and F-measure of 84 percent. The F-measure is defined as the harmonic mean of precision and recall, which can be calculated from the contents of a confusion matrix. This research does not only show the purpose of examining different machine learning methods, but also examines if churn can be predicted accurately. A description of the different methods used in this study will be provided in the next chapter.
3. Research design
3.1 Description
Data from GfK concerning the health insurance sector in the Netherlands will be used in this research. The dataset consists of switching behavior shown by 6445 households in a period of four months. During these four months the contacts within the household have the possibility to switch to another health insurer. The same contact within the different households participated in the questionnaire each month.
The dataset provides the information needed to conduct analyses, but the data required to be cleaned and manipulated in order to be ready for usage. The participants have rarely participated in the panel for more than a month, meaning the data would be more useful without dynamics. In the case that one participant had observations for multiple months, most of the time all observations aside from one entry would be empty. Sometimes multiple observations over multiple months were found, but the information was the same over the months. This means, although it was assumed beforehand that the data incorporated dynamics, that dynamics could not be used in the models for this study. Firstly, the data has been transformed from monthly data into customer-level data, using only the variables needed in this research. The variables which were used for the analyses have been recoded, and the resulting variables will be described in paragraph 3.2.
22 been used to perform multiple imputation using predictive mean matching, with the usage of the predictor matrix.
3.2 Measurement of variables 3.2.1. Measuring the predictors
Specific variables need to be created first from the questionnaire results that the households have participated in. Factor analysis has been used to create factors out of the different questions concerning perceived price of the premium and customer dissatisfaction. Cronbach’s Alpha will be used to analyze whether the factors are fit for further use.
Appendix A describes the variables that have been used in the factor analysis. There are some variables that can be assumed to be relevant to the desired factor of price and also variables that can be assumed to relevant to the desired factor of dissatisfaction. Despite the assumption, all the variables that were part of the same main question (SX7) have been added to the factor analysis, allowing for the exclusion of non-relevant variables when deemed necessary by the results of the analysis.
23 Cronbach’s alpha will show if the chosen variables are definite representations for the factors. For composite reliabilities, values greater than 0.6 are considered desirable (Bagozzi & Yi, 1988). For factor 1 Cronbach’s alpha is 0.9, with a significance of 0.002. The reliability will decrease if one of these items is dropped. For factor 2 Cronbach’s alpha is 0.81, with a significance of 0.0044. The reliability will decrease if one of these items is dropped. The factors can be created:
Factor price = (SX7A + SX7B + SX7C + SX7E + SX7G) / 5 Factor dissatisfaction = (SX7D + SX7K + SX7M) / 3
Originally the variables were ordinal with 4 levels, with a value of 1 meaning that the variable plays a high role in churn, and 4 meaning that the variable does not play a role. These values have been reversed to finalize the creation of the factors. The meaning of the different variable names can be found in appendix A.
3.2.2. Measuring churn
The aim is to build a model that is able to predict which customers are most likely to churn. This will be defined as:
Y𝑖 = {1,0, if customer 𝑖 churnsif customer 𝑖 does not churn
It is found that 11.5 percent of the respondents in the dataset have churned.
3.2.3 Measuring the moderators
24 The usage of a collective scheme is a binary variable and is ready for usage. A change in the personal situation cannot be derived from the data, as no changes are found in the time duration of four months. A question has been asked to the respondents concerning a change in personal situation. If the respondents signal the change has an important role in their decision making, it is assumed that a change in personal situation has occurred. Therefore, it can be used for further analysis. The same will be done for a change in health. Both indicators will need to be converted into binary values for further usage (important vs. not important).
3.2.4. Measuring the mediator
Churn intention can be measured by checking if the participants are planning on changing their health insurance during the period in which they are able to do so. Variable SX4 in the dataset asks this question to the respondents with a 3-scale answer model (“Yes”, “Maybe” and “No”). This variable does not state that someone will actually churn, as it might also mean that a different insurance will be enclosed at the same insurer.
Where earlier approaches to the analysis of mediation relied on some form of structural equation modeling, these methods were not derived from a formal framework for causal inference and did not permit sensitivity analyses with respect to key identification assumptions. Also, earlier methods were difficult to correctly extend to non-linear models such as those with binary outcome variables, as is the case in this study. R has a “mediation” package which implements the procedures described in Imai, Keele, and Tingley (2010), allowing for the conduction of sensitivity analyses and furthermore covering several common statistical models that handle binary dependent variables.
25 By performing a causal mediation analysis with non-parametric bootstrapping, the indirect effect can be measured. In conclusion, this framework is best suited for this study because of the ability to accommodate nonlinear relationships, continuous and discrete mediators, and various types of outcome variables. Therefore, a causal mediation analysis with non-parametric bootstrapping will be used to investigate the mediating effect of churn intention on customer churn.
3.2.5. List of variables
Variable name Level Description
Churn 0
1
No churn Churn Perceived price factor
Dissatisfaction factor 1 2 3 4 No role Small role Important role Very important role Age Household size Continuous 18-90 1-13 Income Education 1 2 3 Low Medium High Collective scheme 0 1 No Yes
Health insurer Insurer CZ
Insurer IZA/IZZ Insurer Menzis Insurer VGZ Insurer Achmea Remaining Insurers “Remaining insurers”
embodies all of the insurers outside of the five largest ones which have their own factor level.
Change in personal situation Change in health situation
26
3.3 Models
The base model is a logit model, with the mediating effect estimated by standard regression.
Logit model
𝑷[𝒀𝒊 = 𝟏] = 𝚲(𝜶 + 𝒙𝒊′𝜷)
Yi is the dependent variable describing churn for customer i
α is the coefficient on the constant term
xi is a vector of the independent variable, covariates and moderators, including the mediating
variable for customer i
β is the rate of change in the “log odds” as xi changes.
Multiple models are made according to the hypotheses. The base model contains the factor of perceived price, the factor of dissatisfaction, the demographic variables and the mediating effect. In order to add the mediation effect to the models, for each hypothesis two separate models are fitted. These fitted objects comprise the main inputs to the “mediate” function in R, which computes the estimated average causal mediation effects (ACME), the average direct effects (ADE) and other quantities of interest under these models, of which the proportion of the independent variable that is mediated is the most interesting one. Identifying the proportion that is mediated through churn intention is the goal of this analysis, along with testing the significance of the indirect effect. To test the hypotheses concerning the moderating effects, for each hypothesis the moderating effect is added to the base model. The model determining P[Yi = 1] can be used for further analysis of the direct effect and other predictors.
Standard regression for the mediator
𝑴𝒊= 𝜹 + 𝒛𝒊𝜸 + 𝜺𝒊
Mi is the mediating variable describing churn intention for customer i
𝜹 is the coefficient on the constant term
zi is a vector of independent variables, covariates and moderators for customer i
27 The specific model for hypothesis 1 will be described in this chapter, and the other formulas for the rest of the hypotheses can be found in appendix B.
Model Specification
Yi = Perceived price of the premium for customer i
Pi = Perceived price of the premium for customer i
DSi = Customer dissatisfaction for customer i
Mi = Mediating effect churn intention for customer i
AGEi = Age of customer i
INCi = Factor income for customer i
HHSi = Household size for customer i
EDUi = Factor education for customer i
εit = Error term
𝑷[𝒀𝒊 = 𝟏] = 𝚲(𝜶 + 𝜷𝟏𝑷′𝒊+ 𝜷𝟐𝑫𝑺𝒊′+ 𝜷𝟑𝑨𝑮𝑬𝒊′+ 𝜷𝟒𝑰𝑵𝑪𝒊′+ 𝜷𝟓𝑯𝑯𝑺𝒊′+ 𝜷𝟔𝑬𝑫𝑼𝒊′+ 𝜷𝟕𝑴𝒊′) 𝑴𝒊 = 𝜹 + 𝜸𝟏𝑷𝒊+ 𝜸𝟐𝑫𝑺𝒊+ 𝜸𝟑𝑨𝑮𝑬𝒊+ 𝜸𝟒𝑰𝑵𝑪𝒊+ 𝜸𝟓𝑯𝑯𝑺𝒊+ 𝜸𝟔𝑬𝑫𝑼𝒊+ 𝜺𝒊
The first equation describes H1, namely the effect of price perception on churn. The second equation describes H1’, namely the effect of the mediator, churn intention, on churn.
3.4 Steps to reject the null-hypotheses and build the model with the best performance
In this paragraph the steps are shown that will be taken to reject the null-hypotheses. The first hypothesis will be used as an example. As a reminder, this hypothesis is:
H1: Negative perceived price of the premium has a positive influence on customer churn This hypothesis is accompanied by the hypothesis describing the effect of the mediator ‘churn intention’:
28 The effect in H1 will be estimated by means of a logistic regression, using the logit model described in the previous paragraph. The results of the logistic regressions can be found in chapter 4.2. The effect in H1’ will be estimated by means of a linear regression, using the formula described in the previous chapter. By using the package ‘mediation’ in R, these two separate formulas can be combined into one formula. This allows us to estimate the indirect effect between the predictors and the mediating variable, which is done in chapter 4.3. One by one the different predictors can be defined in the code, to estimate the indirect effect. For H1’ the variable describing price perception will be entered as specific variable of which the indirect effect needs to be estimated.
The results of the logistic regression and the measurement of the indirect effect will be used to describe whether the null-hypothesis has been rejected. The usefulness of this method lies in the ability to create one model, which includes all the results needed for interpretation purposes concerning the different hypotheses. Furthermore, the estimates of the indirect effect can be measured, which is important when the biggest driver in predicting churn will be identified. This method will be used to interpret all the hypotheses which are accommodated by the different models.
The quality of the different models will be determined through different measures (e.g. AIC, BIC, Log-likelihood). The dataset will also be split into a set in which 50 percent of the participants churned and 50 percent stayed at their current insurer, to check whether this increases the performance. Because the current conceptual model is complicated, it is also useful to check whether the model can be simplified by removing churn from the model and using churn intention as the dependent variable. Eventually the model with the best performance will be used to determine the biggest predictor of churn, and furthermore used to develop different machine learning algorithms.
3.5 Prediction of churn
3.5.1. Determining the hit rate of the models
29
3.5.2 Machine learning techniques
Table 3 describes the methods that will be used to predict churn.
Methods:
1. Logistic regression 2. Support Vector Machine 3. Neural Network
4. Decision Tree 4.1. Regular method
4.2. Bootstrap aggregating (i.e. Bagging) 4.3. Boosting
4.4. Random Forest 5. Naïve Bayes
The logit-model appears to be the most popular model in terms of predicting churn, which is probably because the model is relatively simple, shows good performance, is robust, and the parameter estimates are interpretable in terms of odds ratios (Günther et al., 2014). A drawback of the model is that it assumes linear relationship between the logit and explanatory variables, which is not always the case. When this is not the case, information is lost and the conclusions might not be valid. Support Vector Machines (SVMs), used for classification and regression analysis, are supervised learning models with associated learning algorithms that analyze data and recognize patterns (Vafeiadis et al., 2015).
30 bagged trees. The difference between bagging and random forests is that in building a random forest, at each split in the tree, the algorithm cannot consider a majority of the available predictors. For example, if there is a very strong predictor in the data set, along with an amount of other moderately strong predictors, then in the collection of bagged trees most (or all) of the trees will use this strong predictor in the top split. The result is that bagged trees will be highly correlated, whereas random forests force each split to consider only a subset of the predictors, overcoming the problem that bagging faces (James et al, 2013).
A Naïve Bayes classifier assumes that the presence (or absence) of a particular feature of a class, such as customer churn, is unrelated to the presence (or absence) of any other feature (Vafeiadis et al., 2015). A neural network delivers low classification error rates even when the number of data is not big enough or when significant data noise is present (Nagappan & Ball, 2007).
Multiple algorithms are used, because each algorithm has their different strengths. For instance, the choice can be made between a quick method, or a more accurate method. Random forest, SVM, Boosted Decision Trees and Neural Networks are generally more accurate methods, whereas a regular decision tree or a logistic regression are generally quicker methods. If the classification is still not explainable with the current methods then Naïve Bayes could be a good method to follow up with.
31
4. Results
In this chapter the results of the different analyses are shown. First the correlations between the variables that have been described in the previous chapter will be analyzed, after which the logistic regressions will be performed on the different models. Non-parametric bootstrapping procedures are used to obtain a distribution of causal mediation effects. The highest performed model will be chosen and used to develop machine learning algorithms.
4.1 Correlation table
Table 4 shows the correlations between the different variables that have been described in the previous chapter. This is the first step towards testing the validity of the hypotheses, although correlation does not imply causation, which will be analyzed in chapter 4.2.
Table 4 – Correlation table
32 personal situation (r = .47, p = 0). It is worth noting that the correlation between churn intention and churn is positively weak (r = .36, p = 0), which indicates higher churn intention does not ensure customer churn.
4.2 Logistic regression
Logistic regressions have been performed in order to determine the strength of the different models. Logistic regressions allow for the prediction of a binary outcome variable from continuous predictor variables. Besides, the logarithmic transformation on the outcome variable allows for modelling a nonlinear relationship in a linear way. A final model has also been created, which adds all of the moderators (aside from 2b) to the base model. The equation for the final model can be found in appendix B. The effects described in the hypotheses will also be measured in these logistic regressions. The results of the logistic regression concerning all hypotheses can be found in table 5. All model numbers conform to the same hypotheses numbers.
Churn intention and age have a significant effect on churn in all models. Age has a negative effect on churn, and churn intention has a positive effect on churn.
The factor for perceived price is significant in model 1 (b = .197, p < .10), meaning that H1 can be accepted based this model. This means that an increase in perceived price will lead to an increase in the probability of observing churn. Secondly, the moderating effect described in H2a has been analyzed by calculating an interaction term, which is not significant (b = -.007, p > .10) based on model 2.
34 The moderating effect described in model 2c, concerning the collective scheme, is not significant (b = -.024, p >.10). The moderating effect described in model 2d, concerning the change in health, is not significant (b = -.023, p >.10), and the same is the case for model 2e (b = -.134, p >.10), describing the change in personal situation.
4.3 Mediation analysis
Bootstrapping was used to obtain a distribution of causal mediation effects. This is a non-parametric method based on resampling with replacement which is done 500 times in this study (Bollen & Stine, 1990; Shrout & Bolger, 2002). For each of these samples the indirect effect is computed and a sampling distribution is empirically generated. Table 6 contains the result of the mediation effect of perceived price through churn intention in the base model. Both the direct and indirect effect of perceived price on churn is significant, and 65.9 percent of the effect of perceived price on churn is mediated by churn intention. Therefore, H1’ is approved.
The tables for the other mediator hypotheses can be found in appendix C. The relationship between perceived price and churn is not significantly mediated by the different health insurers, meaning H2b’ is disapproved. The relationship between perceived price and churn is significantly mediated by the moderating effect described in H2a’, H2c’, H2d’ and H2e’. All mediation effects can be seen in table 7.
Table 6 – Causal mediation analysis
Causal Mediation Analysis
Nonparametric Bootstrap Confidence Intervals with the Percentile Method Estimate 95% CI Lower 95% CI Upper p-value Total Effect 0.02170 0.01987 0.02352 0 ACME (average) 0.01430 0.01135 0.01814 0 ADE (average) 0.00741 0.00364 0.01020 0 Prop. Mediated (average) 0.65878 0.53520 0.82883 0
35
Independent variable ACME Proportion Mediated
H1’ Price 0.01430*** 65.9%***
H2a’ Price : Dissatisfaction -0.00863*** 74.0%
H2c’ Price : Collective -0.00856*** 46.7%*
H2d’ Price : Health -0.011961*** 95.2%
H2e’ Price : Personal -0.01173*** 70.2%
Table 7 – Mediation effects ***p < .01; **p < .05; *p < .1 H2a’, H2c’, H2d’ and H2e’ are all approved.
4.4 Comparing the models and determining the best model
All of the models can be validated using different diagnostics, and based on these diagnostics the best model can be determined. Table 8 shows an overview of these different diagnostics, and allow for the comparison of the different models.
Model 1 Price Model 2a Dissatisf Model 2b Insurer Model 2c Collective Model2d Health Model2e Personal Final Model N 6,445 6,445 4,285 6,445 6,445 6,445 6,445 Log Likelihood -1931.56 -1931.55 -947.24 -1928.82 -1931.48 -1930.72 -1927.69 AIC 3883.123 3885.110 1934.480 3881.643 3886.959 3885.449 3889.375 BIC 3950.833 3959.591 2061.737 3962.895 3966.702 3968.212 4004.483 McFadden R2 0.162 0.162 0.232 0.163 0.162 0.163 0.164 Hitrate 0.5 0.895 0.895 0.926 0.895 0.895 0.895 0.895 Hitrate optimal 0.712 0.712 0.782 0.709 0.694 0.712 0.704 Hitrate naive 0.796 0.796 0.846 0.796 0.796 0.796 0.796
Table 8 – Validation diagnostics Findings:
Model 2b (Insurer) cannot be compared to the other models.
According to the LL, Model 2c (Collective) and the Final Model are the best
36
BIC penalizes LL while adding more explanatory variables. Therefore, it is a good way to see if adding additional explanatory variables would significantly improve the fit or not
According to the BIC, model 1 is the best, but also has fewer explanatory variables. Models 2c, d and e can be compared, and model 2c (Collective) is the best according to this criterion.
The McFadden R2 is almost equal across models
According to the optimal hitrate, models 1, 2a and 2e are the best fitted.
Overall, Model 2b (Insurer) appears to be the best performing model. The issue with this model is that a large amount of the sample size is not used compared to the other models. The amount of observations is already relatively low, and for churn prediction purposes it is not recommended to work with a smaller sample size as the dataset will be split during this process. By taking this into account, Model 2c (Collective) appears the overall best model, and this model will be used for machine learning purposes. Because model 2a also performs well according to the LL, the moderating effect of dissatisfaction has been added to Model 2c, but this resulted in a worse model. Therefore model 2c will not be changed further.
In order for the model to efficiently outperform the naïve model in terms of hit rate, the dataset has been split towards a 50/50 churn/no-churn setting. The naïve model is now outperformed by the split model in terms of hit rate, as can be seen in table 9. The main difference between the regular and split model (table 10) is that the constant, age and household size are not significant in the split model. The factor for price is moderately in the split model, as opposed to being significant in the regular model.
Regular Split N 6,445 1,448 McFadden R2 0.163 0.189 Hitrate 0.5 0.895 0.722 Hitrate optimal 0.709 0.722 Hitrate naive 0.796 0.5
37 Although the hit rate of the split model is better, the regular model will be used for further analysis as the main effect is significant. Furthermore, the amount of observations in the split dataset is low compared to the regular dataset and the 50/50 setting of the split dataset is not an accurate representation of the actual churn/no-churn distribution in the health insurance sector.
The model used in this study is complicated, with the addition of both mediator and moderators. To check whether this is necessary, a model has been developed in which churn intention replaces churn as the output variable. The variable describing whether someone actually churns has been removed from the equation. The results have been added to appendix D. The performance of both models can be compared in table 11, in which the extensive model has churn as dependent variable, and the simplified model has churn intention as the dependent variable. Churn model Churn intention model Log likelihood -1928.82 -4985.729 BIC 3962.895 10074.71 AIC 3881.643 9993.457 McFadden R2 0.163 0.087
It is noticeable that the churn model performs at least two times better than the churn intention model. Simplifying the model would damage the quality of the research, thus the extensive model with churn intention as mediator will continue to be used. Furthermore, the correlation table in paragraph 4.1 described a low correlation between churn intention and churn, thus already suggesting that both variables do not describe the exact same phenomenon.
The model has been tested for multicollinearity through interpretation of the variance inflation scores (VIF). In this test the VIF scores are compared to the limit of five (Leeflang, Bijmolt, Pauwels & Wieringa, 2015). All variables have VIF scores lower than five, meaning that there is no multicollinearity in the model.
38
4.5 Machine learning
The dataset has been split into training and testing (80/20). If the assumptions are more or less correct then the data of today is a reasonable representation of the expected data in the future, which is why holding back part of the current data for testing is a fair approximation to future testing. The top decile lift is calculated by using a model to predict churn, using p(churn) to rank all the customers from high to low risk. Customers are divided into ten groups, in which group 1 has the highest p.
𝑇𝐷𝐿 = 𝑎𝑐𝑡𝑢𝑎𝑙 𝑐ℎ𝑢𝑟𝑛 𝑟𝑎𝑡𝑒 𝑜𝑓 𝑔𝑟𝑜𝑢𝑝 1
𝑎𝑐𝑡𝑢𝑎𝑙 𝑜𝑣𝑒𝑟𝑎𝑙𝑙 𝑐ℎ𝑢𝑟𝑛 𝑟𝑎𝑡𝑒 ∗ 100%
The GINI coefficient takes the area between the curve and the diagonal line (A) and the area above the curve (B).
𝐺𝐼𝑁𝐼 = 𝐴
𝐴 + 𝐵
Both the TDL and the GINI coefficient are widely used in machine learning practices to compare the different algorithms. The higher the GINI coefficient, the better the model. Where GINI has a focus on overall performance, TDL focuses on predicting churners. In terms of GINI, the Logit model, the Boosted Decision Tree and the Neural Network have the best performance as they all have a GINI of 0.6 or higher. The calculation time of the Naïve Bayes highly outperforms the other algorithms, which all have comparable calculation times. All models have the same correct prediction percentage.
40
5. Conclusion
The main goal of the research was to study if price was the biggest driver in predicting customer churn in the health insurance sector, and build a model which is able to accurately predict churn. The main hypothesis described that a negative price perception has a positive effect on churn. The factor for customer dissatisfaction was included in the model to allow for the comparison of the estimates of both drivers of customer churn. To elaborate on the main hypothesis, multiple interaction effects on the main effect have been analyzed. These interaction effects are: customer dissatisfaction, use of a collective scheme, type of insurer, a change in health of the customer and a change in the personal situation of the customer. Different models have been created to account for all different moderators. Furthermore, churn intention has been included in the study as a mediator and demographic variables have included as predictors in the analyses. Data acquired from GfK concerning churning behavior in the health insurance sector has been cleaned and manipulated in order to generate results concerning the hypotheses. Tables 11 and 12 describe the different hypotheses and whether they are supported or not.
Hypothesis Support Comment
H1 Negative perceived price of the premium has a
positive influence on customer churn
Supported
H2a The effect of negative perceived price of the
premium on customer churn is stronger if the level of customer dissatisfaction is higher.
Not supported
H2b The effect of negative perceived price of the
premium on customer churn differs across different health insurers.
Partially supported Achmea and “Remaining Insurer” are supported, as compared to “CZ”
H2c The effect of negative perceived price of the
premium on customer churn is weaker if a collective scheme is involved.
Not supported
H2d
The effect of negative perceived price of the premium on customer churn is stronger if a change in the health of a customer is perceived.
Not supported
41
Hypothesis
Support Comment
H2e
The effect of negative perceived price of the premium on customer churn is stronger if a change in the personal situation of a customer is perceived.
Not supported
H1’ The effect of negative perceived price of the
premium on customer churn is mediated by churn intention.
Supported
H2a’ The effect of negative perceived price of the
premium on customer churn when the customer is dissatisfied is mediated by churn intention
Supported Effect is negative, whereas a positive effect was expected
H2b’ The effect of negative perceived price of the
premium on customer churn across different health insurers is mediated by churn intention
Not supported
H2c’ The effect of negative perceived price of the
premium on customer churn when a collective scheme is involved is mediated by churn intention
Supported
H2d’ The effect of negative perceived price of the
premium on customer churn when a change in the health of the customer is perceived is mediated by churn intention
Supported
H2e’ The effect of negative perceived price of the
premium on customer churn when a change in the personal situation of the customer is perceived is mediated by churn intention
Supported
Table 12 – Hypothesis testing (2/2)
42 effect is negative, where a positive effect was expected (Anderson, 1996). Concerning the different health insurers, the results show that the effect of perceived price on churn weakens when the customer is with Achmea or a smaller insurer, compared to CZ. This means that hypothesis 2b is partially supported by the results, as not all the results concerning the different insurers are significant. Surprisingly, the direct effect of these insurers on churn is positive and significant. This would lead to the assumption that there would be other reasons for customers to churn that have not been explored in this study; otherwise the moderating effect would be positive. The effect of perceived price on churn becomes significantly weaker if the customer is part of a collective scheme, and the effect is fully mediated through churn intention. A change in the personal situation of a customer and a change in the health of a customer both significantly influence perceived price, and this effect is fully mediated through churn intention. These effects are positive, as was expected.
When comparing all the different estimates in the different models, it can be noted that perceived price is the biggest driver for churn in this study. This effect differs across insurers. Furthermore, if the effect of perceived price on churn is mediated through churn intention, the effect differs if the customer is dissatisfied, part of a collective scheme, perceived a change in health, or perceived a change in the personal situation.
The indirect effect of the different drivers on churn through churn intention has been analyzed by performing a causal mediation analysis with non-parametric bootstrapping. The results showed all of the moderating effects, aside from the one describing the different insurers, are mediated through churn intention. The results of the correlation analysis that had been conducted beforehand showed a weak correlation between churn and churn intention, meaning that it would be important to continue with the current model, and not omit the variable for customer churn.
In order to build a machine learning algorithm that can accurately predict churn for further interpretation, the different models have been compared in order to decide on the strongest model. This resulted in model 2c, which includes the moderating effect of the usage of a collective scheme to the main model, being the best model. The differences between the performances of the different models are small, but one model needed to be chosen for prediction purposes.
43 observations in the model with different insurers are missing. Therefore, it is difficult to state in this study whether the model with the different insurers is better than the model with the collective scheme.
Machine learning algorithms have been developed. The algorithms for the logistic regression, neural network and the boosted decision tree outperform the other algorithms in terms of model performance.
Scientific implications
In conjunction with the study from Lim et al. (2006), it has been established economic value is a cause of customer churn; a construct which has been shaped in the form of perceived price in this study. Furthermore, some demographics factors are also a cause of customer churn, as has been supported by previous studies (Ahn et al., 2006; Burez & Van den Poel, 2009; Ferreira et al., 2004; Lee et al., 2011; Lemmens & Croux, 2006). Customer dissatisfaction as a direct cause of customer churn could not be established, contrary to previous findings (Ahn et al., 2006; Lim et al., 2006). The negative effect of customer dissatisfaction as a moderator on the effect of perceived price on churn is contrary to previous research by Anderson (1996), where a positive effect was found. Anderson (1996) notes that firms providing higher customer satisfaction will not necessarily have customers with greater price tolerance, and follows by stating that although customer satisfaction is found to be higher for firms in more competitive categories (Anderson, 1994), price tolerance may be lower. Recall that Dutch customers are allowed to switch health insurer once a year, thus encouraging competition (Maarse et al., 2016). This provides an explanation for the negative effect of customer dissatisfaction as a moderator on the effect of perceived price on churn, mediated by churn intention.
