New developments in churn prediction models:

(1)

New developments in churn prediction

models:

testing the impact of social influence, historical data and

customer value in the health insurance market

(2)

New developments in churn prediction models:

testing the impact of social influence, historical data and customer value

in the health insurance market

Bas Cuperus

Meeuwerderweg 119a

9724 ER Groningen

Tel: +31(0) 614136448

E-mail:

b.cuperus@student.rug.nl

Student number: S1994883

Master Thesis Marketing Intelligence

January 16, 2017

Supervisors:

First supervisor: dr. ir. M.J. Gijsenberg

Second supervisor: dr. J.T. Bouma

University of Groningen

Faculty of Economics and Business

Department of Marketing

(3)

3

Management Summary

Customers defect more and more, which is a main concern for firms that rely on customers as their principal asset. Especially in markets that do not significantly grow anymore,

customer retention receives a large amount of attention by operating firms and scientists. In order to manage churn, firms predict which customers are most likely to churn using

statistical analyses on historical data. The customers with the highest churn probabilities, are then targeted for retention campaigns in order to keep them on board.

Predictive variables are (almost) always based on some socio demographic/economic characteristics as well as on the customers’ relationship with the firm. However, some developments found their way into research, adding value to the churn model literature with new approaches. This paper deals with three of these important developments: first, the impact of social factors on churn decisions. Second, the use of longitudinal data, so that patterns over time can be found. Third, adding an understanding of the heterogeneity in value customers represent (i.e. the value at risk) in retention program targeting, so that programs are created in order to create maximal profitability. The current study is unique, valuable and important because it is the first to combine the above mentioned findings and assumptions into one model, in order to see the (relative) improvement of prediction and information.

Adding a social influence variable and historical data to the model, resulted in highly significant, strong beta’s which influenced the predictive power of the churn model in a positive manner. It would be valuable for managers to keep in mind that individual

(4)

4

Preface

This thesis marks the end of my time as a student as well as the start of a new phase of me as a professional. I enjoyed writing this paper and I am very grateful for the help I received during the process.

I would like to thank my supervisor Maarten Gijsenberg for his help, inspiration and

feedback, and supervisor Jelle Bouma for his help related to my internship and thesis. I also would like to thank the Marketing Intelligence department at Univé for the opportunity to work with them, their trust and all the insights I gained during my internship. Last but not least, I would like to thank my family, friends, fellow students and roommates for their support and help during my time at the university.

Bas Cuperus Groningen

(5)

5

Content

Management Summary ... 3 Preface ... 4 1. Introduction ... 7 2. Theoretical Framework ... 9 2.1 Churn management ... 9 2.1.1 Prediction methods ... 9

2.2 Common drivers of churn ~ the benchmark model ... 10

2.2.1 Relational Characteristics ... 10

2.2.2 Socio Demographics ... 11

2.3 Further optimizing campaign/retention decisions ... 11

2.3.1 Improve predictive model power: the networked customer... 12

2.3.2 Improve predictive model power: history matters ... 13

2.3.3 Improve profitability of decision: capture customer value ... 13

2.4 Conceptual model ... 14

3. Research design ... 14

3.1 Research setting and data overview ... 15

3.2 The benchmark model ... 16

3.2.1 The benchmark model ~ the estimation technique ... 16

3.3 Sample ... 16

3.4 Judging on model performance with the top decile lift ... 17

3.5 The social influence variable ... 17

3.6 Longitudinal data ... 18

3.7 Customer value ... 18

4. Results ... 19

4.1 the benchmark model in a balanced sample ... 19

4.1.1 Benchmark model performance ~ full dataset ... 21

(6)

6 4.2.1 Model performance ~ Adding the social influence variable and the longitudinal data

... 26

4.3 Model Performance: paying attention to customer value ... 26

5. Conclusions and discussion ... 27

5.1 Theoretical assumptions in the benchmark model ... 27

5.2 Research question 1: To what extent does the addition of a (basic) network variable add value to the predictive power of the churn model (compared to the benchmark model)? ... 27

5.3 Research question 2: To what extent does historical data add value to the predictive power of the churn model (compared to the cross-sectional benchmark model)? ... 28

5.4 Research question 3: To what extent do the calculated churn probabilities support retention targeting better when adjusted for customer value? ... 28

5.5 Managerial implications ... 28

5.6 Limitations and ideas for further research ... 29

Appendix 1: the initial independent variables of the benchmark model ... 31

Appendix 2: the extra variables added to the total dataset... 33

Appendix 3 the cleaned benchmark model ~ validation ... 34

Appendix 4 The cleaned benchmark model betas ... 35

Appendix 5 the full benchmark model ~ validation ... 39

Appendix 6: the uncleaned model with social influence and longitudinal data ... 40

Appendix 7: the cleaned model with social influence and longitudinal data ... 41

Appendix 8: parameters cleaned model (Social and Longitudinal) ... 42

(7)

7

1. Introduction

Customers defect more and more, which is a main concern for firms that rely on customers as their principal asset. The growing amount of customers that defect from a service

company is the result of growing awareness of switching opportunities and the weakening of switching barriers due to governmental deregulations and growing market transparency (Holtrop et al., 2016). Customer loyalty is not to be taken for granted and reported churn rates confirm this.

These reported churn rates in the literature can be as high as 63% a year (Holtrop et al., 2016). The financial consequences of churn, defined as the end of the total revenue stream from a specific customer (Braun and Schweidel, 2011) can be dramatic (Lemmens and Gupta, 2013). First , the normal revenue stream has been cut off. Second, all opportunities for up and cross selling for this specific customer are lost (Risselada, Verhoef and Bijmolt, 2010). Third, acquiring a new customer (to replace the churned customer) is far more expensive than retaining the current customer (Verbeke et al., 2012).

Especially in markets that do not significantly grow anymore, customer retention receives a large amount of attention by operating firms and scientists (Verbeke et al., 2012; Risselada, Verhoef and Bijmolt, 2010; Neslin et al., 2006). This can be explained by the fact that there are no new customers on the market; new customers obtained through acquisition are the former customers of the competitor. One man’s meat is another man’s poison is the rule at this type of markets. Since competition revolves around the (most profitable) customers and acquisition costs continue to rise (Verbeke et al., 2012; Lemmens and Gupta, 2013),

retention actions are becoming more important for firms. Customer retention is an important driver for customer equity (the sum of all customer lifetime values) and firm value, because it affects expected incoming, future cash flows and profitability.

(8)

8 This paper deals with three of these important developments: first, the impact of social factors on churn decisions (Haenlein, 2013). Second, the use of longitudinal data (Ascarza and Hardie, 2013), so that patterns over time can be found. Third, adding an understanding of the heterogeneity in value customers represent (i.e. the value at risk) in retention

program targeting, so that programs are created in order to create maximal profitability (Lemmens and Gupta, 2013). The first two developments are important because they show significant and substantial improvements in the predictive power of churn models (Haenlein, 2013; Nitzan and Libai, 2011; Ascarza and Hardie, 2013). The third development is

important because it adds value to the idea that retention programs are there to maximize the customer equity and profitability, rather than only keeping every single customer with the firm (Lemmens and Gupta, 2013). A new way of modeling with an adjustment for

customer value, proposed by Lemmens and Gupta (2013), resulted in an average increase of 115% in profits.

The current study is unique, valuable and important because it is the first to combine the above mentioned findings and assumptions into one model, in order to see the (relative) improvement of prediction and information. The contribution to science and to managerial knowledge is twofold. First of all, the network, longitudinal data and customer value constructs which have not received the necessary attention so far, are researched again in order to contribute to the general acceptance of these constructs. Secondly, the constructs were always researched separately. Now they are combined into one paper. This means that the relative importance of the constructs is becoming more visible and clear.

The main research question therefore is:

In what way does retention targeting change when recent developments from churn research are taken into account?

From this main question, three relevant sub questions emerge:

- To what extent does the addition of a (basic) network variable add value to the

predictive power of the churn model (compared to the benchmark model)?

- To what extent does historical data add value to the predictive power of the churn

model (compared to the cross-sectional benchmark model)?

- To what extent do the calculated churn probabilities support retention targeting

better when adjusted for customer value?

(9)

9 can be found. After the theoretical fundament follows a chapter in which is explained how the research in this paper is operationalized. The fourth chapter shows the results of the research done for this paper and in chapter five the results will be discussed in both a practical managerial way as well as in a scientific way.

2. Theoretical Framework

2.1 Churn management

Firms strive for establishing long-term relationships and maximizing customer equity by managing churn. Especially in markets where services are sold on a contractual basis, churn is the defection of a relationship instead of just one product less sold (Risselada, Verhoef & Bijmolt, 2010). In order to manage churn, firms predict which customers are most likely to churn using statistical analyses on historical data. The next paragraph deals with the different techniques that are used for these predictions. The customers with the highest churn probabilities, are then targeted for retention campaigns in order to keep them on board. Campaigns can consist of incentives like discounts and special offers to increase loyalty (Holtrop et al., 2016).

2.1.1 Prediction methods

There are different methods in the literature to predict churn. The most common and most often used methods by researchers and practitioners are logistic regression and

(10)

10

2.2 Common drivers of churn ~ the benchmark model

The predictive variables in the churn literature can mostly be divided in two or three groups: relational characteristics and socio demographic characteristics and/or socio economic characteristics (Holtrop et al., 2016; Risselada, Verhoef & Bijmolt, 2010). The relational characteristics consist of the variables that characterize the relationship of a customer with the firm. Out of the mentioned characteristics, these relational characteristics seem to be most important (Holtrop et al., 2016; Risselada, Verhoef and Bijmolt, 2010). This importance is based on the relative large number of appearances in the different decision trees in these articles. This paragraph shows all important drivers of churn included in the benchmark model.

2.2.1 Relational Characteristics

Relationship length is likely to have a negative relationship with churn probabilities.

Schweider, Fader and Bradlow (2008) specify this as negative duration dependence; i.e. as relationships with subscribers last longer, (service) churn probabilities decrease. Anderson and Weitz (1989) have actually found this same phenomenon; relations that lasted longer, tend to be more stable in later phases. Bolton (1998) confirms and explains this

phenomenon: customers who have more experience with the firm (in terms of time) weigh new information less heavily and prior cumulative satisfaction more heavily.

Lemon, White and Winer (2002) found strong empirical evidence for a relation between the expectations of future product/service use and retention probability. If customers expect to make ‘more’ use of their products/services in the future, it is likely that those customers favor the more comprehensive products/services in the present. Therefore product ranking information is included for prediction purposes, with the extended products ranking higher than the simple products. In other words, customers with a more simple product portfolio are more likely to churn.

Next to the normal and additional products/services, some firms also offer ‘unrelated products’. In terms of the telecommunication market, bundling products for customer retention purposes is called triple play/quadruple play (Ferguson and Brohaugh, 2008). Selling television next to internet is an example of this strategy. In order to see whether product bundling is appropriate for retention, these unrelated products are taken into account in this paper. Risselada, Verhoef and Bijmolt (2010) divide the relationship

(11)

11 on several levels in the research of Risselada, Verhoef and Bijmolt (2010). The indicator length is described as the relationship length (beta’s up to -.3625 in the insurance sector) and the indicator depth is defined by two variables: indication collectively insured (beta’s up to -.4183) and insurance package type (beta’s up to -1.103).

Renewal of the contract is an important variable for churn predictions (Ascarza and Hardie, 2013). Since churn in the past can be seen as a negative renewal, the predictive value of this variable shall be addressed as well.

2.2.2 Socio Demographics

Age is a variable that is expected to be negatively related to churn probabilities; the older a person gets, the less likely this person is about to churn (Haenlein, 2013). Risselada, Verhoef and Bijmolt (2010) found in their health insurance data a quite similar pattern; age was a splitting variable for all of the created decision trees and was significant in two out the three created logit models. The signs for age were always negative. The fact that there are strong signs for a negative relation between age and churn probabilities, could possibly be

explained by the fact that elder persons (i.e. 60+) have more stable preferences than the young consumers, which is positive for repurchasing (Mittal and Kamakura, 2001). The indication whether an individual moved is expected to be positively related to churn probabilities. The product/service provider can be more or less embedded in another environment. In the article of Risselada, Verhoef and Bijmolt (2010) this variable was significant in two out of three years in the logit models they used. Surprisingly it had (relatively strong) negative signs. This could be due the fact that the variable does not present a change of address in the last year, but during the total length of the relationship.

Other personal characteristics like gender, education and urbanization (number of inhabitants of the customers residence) are also found to have an influence on churn decisions (Nitzan and Libai, 2011; Knox and van Oest, 2007; Neslin et al., 2006). The

importance of the constructs are doubtful as they are not that often mentioned as the earlier mentioned personal characteristics. Also their effects (beta’s) are relatively weak compared to the other personal characteristics. But since gender, education and urbanization are mentioned in literature, these variables are taken into account as drivers of churn in this paper in order to be as complete as possible.

2.3 Further optimizing campaign/retention decisions

Besides the well-known relational characteristics and socio demographic predictors as

(12)

12 research (Haenlein, 2013; Ascarza and Hardie, 2013; Lemmens and Gupta) showed new developments, which can help researchers and decision makers to optimize their models even further. The social environment, development/evolution of a customer, and value of a customer were taken into account only recently and until now always separately. The

theoretical fundament and justification for testing the recent developments (for the first time together) can be found in the next paragraphs.

2.3.1 Improve predictive model power: the networked customer

In marketing, it is known and widely accepted that customer behavior cannot be fully explained by studying the individual object separated from its environment. Evidence shows the presence of social influence in product diffusion/adaption setting. Nitzan and Libai (2011) present an overview of papers where social influence is key in these types of settings.

Besides this large span of collected evidence related to the acquisition of customers, even more important evidence is found: quitting behavior is researched in different settings (quitting smoking, defecting from military service and leaving employees) and was the inspiration for Nitzan and Libai (2011) to research social interactions in churn decisions. Nitzan and Libai (2011) found evidence for the influence that social interactions have on churn decisions (in a cellular phone setting). The found effect is comparable in nature and effect to the influence social interactions have on product adoption: the extent decreases over time (i.e. the further in time from the defection event of a member in the direct social network of the focal customer, the lesser the effect on the focal customer) and the likelihood of defection depends on tie strength and the amount of homophily with the defecting

member of the social network as well as the average number of connections this member has.

(13)

13 strong beta (.472). These hazard rates hold up for a period of five weeks after churn (of the social contact).

2.3.2 Improve predictive model power: history matters

Holtrop et al. (2016) acknowledge that in using cross-sectional data, information gathered in the past is ignored. Since parameters of churn models are not per se stable over time

(varying in significance, direction and/or size), the staying power of a cross-sectional

estimated churn model is not that good; if the model is reused again in a subsequent period instead of re-estimated, the predictive power suffers.

Ascarza and Hardie (2013) developed a Hidden Markov Model, which they use to predict future usage and churn. The unique thing that can be done with this model, compared to the traditional cross-sectional estimated models, is to define or capture certain ‘paths to death’. What is meant by these paths of death is the existence of patterns (a journey through different stages) before a customer finally churns. If certain patterns exist, longitudinal/panel data should add value to the predictive performance of a churn model.

The idea of the patterns/paths to death are explored by multiple techniques in the Ascarza and Hardie (2013) paper. The main idea was to predict multiple periods ahead in terms of usage and renewal. Although the logit model, made with longitudinal data performed worse than the cross-sectional one and the Hidden Markov model (the error terms were relatively high, hit rate was worse etc.), the idea that longitudinal data in a logit model adds value is still a starting point in this thesis. This is justifiable due to the fact that Ascarza and Hardie (2013) show the value of longitudinal data, but predict multiple periods ahead instead of one, while the staying power of models are rather low. It is conceivable that methods

perform rather different when making an estimate for just one period ahead instead of more than one period (Risselada, Verhoef and Bijmolt, 2010).

2.3.3 Improve profitability of decision: capture customer value

The aim of churn management should be the maximization of profits (Lemmens and Gupta, 2013). However, current approaches in the field of churn modelling, model the probabilities that customers churn and are therefore only focused on the right classifications (churned or not). The top fraction; the individuals with the highest churn probabilities are then targeted for the retention program. These programs often consist of discounts, personalized or special offers, personalized mailing etc., with the aim to generate more customer loyalty.

(14)

14 represents on an individual level (among other things, which are out of scope for this

research) into account. The authors show that models that may predict churn worse, perform better as input for retention campaigns in terms of profits.

The result accomplished in the Lemmens and Gupta (2013) paper, was an average increase of 115% in profits, while keeping the implementation of the retention campaign the same. The change in size and composition of the target group were the only changes that triggered this increase in profits.

Since the marketing profession is challenged to demonstrate the value it creates by its actions (Srinivasan & Hanssens, 2009), the idea of adding a value component to retention campaigns is a welcome development. Firms steer more and more on customer lifetime value, which is not only a useful metric for tactical decisions, but also a relevant metric to assess the overall firm value (Gupta, Lehmann and Stuart, 2004). The heterogeneity in value individual customers represent, makes that every individual has a different (monetary) influence on the value of the customer base and it is therefore that it is important to account for this heterogeneity while creating churn models.

2.4 Conceptual model

A visualization of the discussed concepts can be found in the form of figure 1.

Figure 1: Conceptual model

3. Research design

As mentioned in the first two chapters of this paper, the influence of social networks and longitudinal data on the predictive power of the model will be researched as well as the influence of customer value on targeting will be researched.

(15)

15 included and what estimation technique will be used. After, the use of balanced samples for estimating the benchmark model will be explained and performance measurement using top-decile lift. After defining the fundament/benchmark, the new developments will be

introduced subsequently.

First the created variable to measure social influence will be explained. Then the method to deal with longitudinal data will be discussed. Finally, the definition of customer value in terms of this research will be presented as well as a method to compare the churn predictions with churn predictions which are adjusted for customer value.

3.1 Research setting and data overview

The research questions mentioned in the introduction will be answered with the help of a dataset provided by a Dutch insurance company. This dataset consists of data from the customer base in health insurances between 01 November 2011 and 01 February 2016: about 1.2 million unique customers. November and February are the two moments a year the dataset gets an update, because a customer can only churn or make adaptions in the product once a year.

The analysis is a prediction of churn in 2015 with the help of the data in 2014. The historical data for research question two will be historical data from the customer who were active in 2014. The full customer base in the year 2014 for healthcare insurances contains about 370.000 unique main insured customers.

(16)

16

3.2 The benchmark model

For the first part, all predictive variables mentioned in paragraph 3.1 will be initially included. After cleaning the benchmark model, some variables will be removed due to the fact that these variables do not contribute enough to the predictive power of the model. The dependent variable (churned) is coded in a binary (dummy : yes/no) way.

3.2.1 The benchmark model ~ the estimation technique

The benchmark model is a logistic regression model. This is a very common technique to estimate churn models in practice. In the research of Neslin et al. (2006) it appeared that 45% of the churn model practitioners (academics as well as firms) use this technique for estimation purposes. It appears to be by far the most popular technique and shows a good short-term predictive performance (Neslin et al., 2006). Donkers, Verhoef, and De Jong (2007) IN Risselada, Verhoef and Bijmolt (2010) even suggest a better performance for logistic regression in the specific (insurance) market. The reason for this popularity, since there are lots of alternatives, lies in the fact that other techniques are far more complicated, without bringing substantial gains in predictive performance with them (Risselada, Verhoef and Bijmolt, 2010). Another very popular method is the use of regression trees due to its ease and its predictive power (Neslin et al., 2006). There is no clear superior method, but the choice for one of the methods should depend on the data structure. If datasets are relatively large (>1.000 records), far from normally distributed and contain a lot of

categorical variables regression trees perform better (Risselada, Verhoef and Bijmolt, 2010). Normally regression trees would thus perform better on the dataset provided by the

insurance firm, due to the large number of records and the presence of a lot of categorical variables, but since there is a serious risk in overfitting the data, a logistic regression model will be estimated. This risk for overfitting is due to the fact that noise and signal are not that easy to separate (Risselada, Verhoef and Bijmolt, 2010). Pseudo R² measures for earlier estimated churn models by the firm were relatively low (see results chapter).

3.3 Sample

In order to make the sample, the size and composition will be discussed in this paragraph. A balanced sample will be used for creating the model, while the 2014 population will be used for validation issues. After the validation, the 2015 population is used to see the predictive power of the model (compared to the benchmark).

(17)

17 population size (around 370.000). Filling in the numbers leaves us with the advice to use a sample size of about 10.000.

Holtrop et al. (2016) refer in their article to prior research when it comes to justifying the composition i.e. the balanced sample: this type of sample performs more reliably without losing efficiency in comparison to proportional samples. This makes sense, as Lemmens and Gupta (2013) explain: non-churners may dominate the statistical analysis and hinder the detection of churn.

3.4 Judging on model performance with the top decile lift

The top decile lift is a technique to validate the success of the churn model, by revealing the difference between the actual churn rate among the ten percent individuals with the highest predicted churn probabilities and the average churn rate. This is done by dividing the

fraction of churners in the top decile by the fraction of churners in the whole set (Lemmens and Croux, 2006). The higher the top decile lift, the better the difference between the top ten percent predicted churners and the average churn in the whole set is visible, i.e. the better the classifier. Another common used validation metric is the Gini Coefficient, which represents the area between the cumulative lift curve and the random prediction. As the Gini coefficient strongly correlates with the top decile lift and the predictive performance in the top decile is key (the other deciles are less important for targeting issues), the top decile lift is the metric which is used for making any decisions for this research (Neslin et al., 2006). Gini coefficients will be reported in order to be as complete as possible.

3.5 The social influence variable

The social influence construct is measured before by Haenlein (2013) and by Nitzan and Libai (2011) in a telecom setting. The influence in their papers was measured by using outgoing call relationships. Since the insurance business does not have data on that kind of objective ties, some modifications were asked in order to come up with a decent solution. The networks that can be found in the health insurances industry is that of multiple insured on one insurance number. The roles the insured can take is the one of main insured or the one co-insured. The co-insured are often related to the main insured in terms of

(18)

18 is measured in this research lies in the influence the co-insured has on the main insured: if a co-insured churns, does the main insured churns next year? The network aspect in this research paper is different and limited compared to earlier mentioned papers in an important aspect: the network is smaller and is only measured within the boundaries of households. Due to the limited size of the networks, the to be found effects can be a lot weaker. But due to the strong links in the network, effects can be stronger as well.

3.6 Longitudinal data

The dataset contains information over multiple years; the variables are measured two times each year. Because of these multiple measurements over time, there is information about changes in customer characteristics (for example: did a customer move) and relational characteristics (for example: did the customer changed his/her product portfolio) over time. This enables to search for paths to death, as mentioned in the theory chapter. Several variables are created to capture the longitudinal effects which relate to collective insurances, the voluntary own risk, the number of policies a customer has over time, the additional health care insurance policies a customer has over time and indications whether the customer moved in the years before. These variables will show the result of changes over the years in the state of the customer on the probability that he/she churns and should help to predict customer churn more in advance. Explicit definitions and the operationalization of these variables can be found in appendix 2.

3.7 Customer value

After the optimal predictive model is prepared, the outcomes (i.e. the probabilities) will be multiplied with the value a customer represents. This value is indicated by the insurance fee paid by the customer, excluding all third party products and all healthcare insurance

products. The reasons for excluding these products are data issues and limited value. There are several issues on the data of third party products. To take them into account would provide heavily distorted and non-realistic outcomes. Data for healthcare insurance fees are not available at all. The second reason to not use these products is their limited value: third party products do not provide any monetary value for the central organization (only for local offices) and due to the way the organization is structured, the margins on healthcare

(19)

19

4. Results

4.1 the benchmark model in a balanced sample

The initial benchmark model, made by the firm, needed to be checked and adjusted. Some tests for model assumptions were made. The cleaned benchmark model is made on the balanced sample (as described in the research method section). All the variables that were included in the dataset are presented and described in appendix 1.

First, all the variables needed to be made numeric in order to search for correlations between independent variables. All categories from categorical variables were numbered. The correlations were searched by using the Spearman Correlation test. Two strong (>.5) correlations were found: introduction discount correlated strongly with relationship length (.-674, p = .000) which makes sense, because introduction discount is for the period of one year to introduce the insurances to the customer (acquisition purpose). Another strong correlation found, is the one between the number of unique policies and car insurance (.769, p = .000), which indicates that customers with multiple (unique) policies often have a car policy. Due to these strong correlations, the number of unique policies variable (categorical in form, see appendix 1) was turned into dummy variables in order to do a regression analysis and obtain VIF (Variance Inflation Factor) scores. With these VIF scores,

multicollinearity issues are researched. The VIF scores for regressing introduction discount and relationship length on each other do not exceed 1,000 and the VIF scores for regressing car insurance with the number of unique policies (three dummies for this variable created) on each other do not exceed 1,091. Due to these low VIF scores, the conclusion is that there are no multicollinearity issues. Leeflang et al. (2015, p.140) describe that multicollinearity becomes a problem when the VIF-scores exceed a value of five.

The second step is to test on information criteria, pseudo r squares and hit rates to see whether there are unnecessary variables included, which do not contribute enough to the predictive power of the model and therefore make the model too complicated. The model is estimated with Backward LR; in every step the least significant variable is excluded. The method provides nine versions of the (nested) model. In the ninth step, the variables: became collectively insured, dental insurance, became main insured, type of residence, education level, fire insurance, co insured young child, co insured old child and co insured newborn were removed. The validation results and the variables included in the final

(20)

20

Model 1 Model 9 (nested version)

Cox and Snell R2 .134 .131

Nagelkerke R2 .179 .174

Hit Rate 64,5 64.4

Hitrate for value of 1 67,2 67,4

Log Likelihood -12142,700 -8037,530

AIC 24417,400 16143,059

AICC 24417,843 16143,178

BIC 24939,030 16411,778

CAIC 25005,030 16445,778

Table 1: model comparison

Due to the better results in information criteria, the beta’s of the nested model will be tested on the full dataset in order to see how the model performs. The input for the full formula can be found below and justification for the used beta’s can be found in Appendix 4.

Variable Name Beta

Intercept 1.233** Age: 80+ -1.384** Age: 65-79 -.599** Age: 50-64 .124 Age: 34-49 .354 Age: 25-34 .630** Age: 18-24 .463* Age: Unknown 0 Churned Before 1.234** Introduction Discount .409** Collectively Insured -.130**

Voluntary Own Risk .183**

Voluntary Own Risk Raised .326*

Addition Health Insurance Policy: Min. Income

-1.648**

(21)

21 Addition Health Insurance Policy: None 0

Addition Health Insurance Policy Dumped .198*

Moved .276** Social Class A .335** Social Class B1 .243** Social Class B2 .201** Social Class C .195** Social Class D 0 Urbanization > 250.000 .242** Urbanization 100.000 – 250.000 .280** Urbanization 50.000 – 100.000 -.021 Urbanization 20.000 – 50.000 .091 Urbanization 10.000 – 20.000 .004 Urbanization 5.000-10.000 -.065 Urbanization < 5.000 0

Number of Unique Policies: Gold -.366** Number of Unique Policies: Silver -.250** Number of Unique Policies: Bronze -.009 Number of Unique Policies: Nickle 0

Car Insurance -.465**

Co Insured Adult .275**

Relationship Length -.539**

Table 2: The beta’s (** significant at a .01 level, * significant at a .05 level)

4.1.1 Benchmark model performance ~ full dataset

(22)

22

Figure 2: TDL of the cleaned benchmark model

4.2 Adding the social influence variable and the longitudinal data to the benchmark

Adding the social influence variable and the longitudinal data in the model resulted in model 10. For the purpose of testing whether historical data added value to predictive modelling, some extra variables were created as presented in Appendix 2. In order to fully understand what happened, an extra table is made in the form of table 5. This table presents all the variables made and their cross-sectional (2014) versions.

Again correlations of all the variables included in model 10 are checked. The Pearson correlation test resulted in four strong (>.5) correlations: relationship length with

introduction discount (-.674, p = .000) and the Number of times voluntary own risk raised since 2011 correlates strongly with the indication voluntary own risk (.611, p = .000), indication voluntary own risk raised (.684, p =.000) and indication voluntary own risk raised 2013 (.663, p = .000). Regressing the correlated variables on each other resulted in VIF scores which did not exceed the value of 1; multicollinearity is not a problem.

Cleaning model 10 the same way as the benchmark model resulted in model 11. Some variables which were created to add value through historical data did not made it to the final model. In table 3 the validation metrics are presented and in appendices 6 and 7 the SPSS output related to this table can be found.

0 10 20 30 40 50 60 70 80 90 100 0 1 2 3 4 5 6 7 8 9 10

Lift Curve

(23)

23

Model 10 Model 11 (nested version)

Cox and Snell R2 .137 .136

Nagelkerke R2 .182 .182

Hit Rate 64,8 64,9

Hitrate for value of 1 66,3 66,3

Log Likelihood -9047,769 -8836,153

AIC 18181,539 17750,306

AICC 18181,728 17750,462

BIC 18521,389 18058,542

CAIC 18564,389 18097,542

Table 3: model comparison

Due to the slightly better results in information criteria, the beta’s of the nested model will be tested on the full dataset in order to see how the model performs. The beta’s for the full formula can be found below in table 4 and all the variables created for the purpose of testing the value of historical data (next to their 2014 versions) can be found in table 5.

Variable Name Beta

Intercept .505 Age: 80+ -.844** Age: 65-79 -.067 Age: 50-64 .644** Age: 34-49 .876** Age: 25-34 1.155** Age: 18-24 .943** Age: Unknown 0 Churned Before .882** Introduction Discount .253* Collectively Insured -.124**

Addition Health Insurance Policy: Min. Income

-1.621**

(24)

24 Addition Health Insurance Policy Dumped .247**

Moved .255** Social Class A .331** Social Class B1 .239** Social Class B2 .199** Social Class C .194** Social Class D 0 Urbanization > 250.000 .251** Urbanization 100.000 – 250.000 .282** Urbanization 50.000 – 100.000 -.008 Urbanization 20.000 – 50.000 .094 Urbanization 10.000 – 20.000 .007 Urbanization 5.000-10.000 -.062 Urbanization < 5.000 0

Number of Unique Policies: Gold -.309** Number of Unique Policies: Silver -.213** Number of Unique Policies: Bronze .026 Number of Unique Policies: Nickle 0

Car Insurance -.452**

Co Insured Adult .265**

Became Collectively Insured ‘12 -.307* Addition Health Insurance Policy Dumped

‘13

.351**

Addition Health Insurance Policy Dumped ‘12

.222*

Moved ‘13 .273**

Relationship Length -.505**

Change in Number of Unique Policies 13-14 -.158**

Social Influence .662**

Number of times Voluntary Own Risk Raised Since 2011

.404**

(25)

25

Total occurrences Sample occurrences Exp. beta (final model) Indication collective 90887 4165 .884

Became collective 4894 226 Non-significant Became collective ‘13 4.055 162 Non-significant Became collective ‘12 4.958 220 .736

Voluntary own risk raised 4435 337 Non-significant Voluntary own risk raised ‘13 954 64 Non-significant Voluntary own risk raised ‘12 4.791 316 Non-significant Number of times voluntary

own risk raised since 2011

10.015 705 1.498

Negative ahi change 9025 624 1.280 Negative ahi change ‘13 14.027 935 1.420 Negative ahi change ‘12 8.626 586 1.249

Moved 27.786 2006 1.291

Moved ’13 13.807 928 1.314

Table 5: Variables and their historical/longitudinal versions (non-significant variables are not included in the model)

There are several variables made on the collectively insured, from which two variables are significant and two not. The indication became collectively insured in 2014 was already removed in the cleaned benchmark model and the indication became collectively insured in 2013 was removed in the final model. It looks like the collective insurances only matter for churn probabilities when the collective situation lasted for a certain time (≥3 years). The variable which indicates whether a person is collectively insured contains 90% of persons which are insured in a collectively manner for more than three years.

The second group of variables contains information about the developments in voluntary own risk. Adding the historical variables (2012 and 2013) made the two indications from the benchmark insignificant. To obtain as most information as possible in a simple manner a last variable was created which indicates how many times the voluntary own risk was raised since 2011. Information from the past has value for predicting the churn probabilities. The third group of variables contains information about removing additional health insurance policies. The beta for removing these policies in 2013 is the strongest for prediction of churn in 2015. This indicates that a relatively large customers give the firm a warning sign before they eventually churn.

(26)

26

4.2.1 Model performance ~ Adding the social influence variable and the longitudinal

data

Scoring out the formula on the full dataset resulted in a Top Decile Lift value of: 2.64 and a Gini Coefficient of: .36. These results are slightly better than the values provided by the benchmark model. Performance improvement thus appears limited. We will discuss this in next chapter. The total lift curve is presented below in figure 3.

Figure 3: TDL of the model with social influence variable and longitudinal data

4.3 Model Performance: paying attention to customer value

After estimating the full model including social influence and historical data, the outcomes (chances) are multiplied by the value a customer represents. In this paragraph a new TDL will be presented as well as the value that the churned customers present for each model (benchmark, including social and longitudinal, adjusted for customer value).

Multiplying the individual outcomes of the model by the fees paid by that customer (=value at risk) results in another top decile. The lift is 1.32; only half of the best predictive model. The main idea of multiplying with customer value is not to add predictive power to the model, but to select a top decile which represents more value to a firm in order to keep that value/fees with the firm. In this, the method succeeded. The churned value the top decile adjusted for customer value contains is: €2.969.069,95 whereas the churned value in the top decile of the social influence and longitudinal model is: €770.412,24 and the churned value of the benchmark model is: €769.037,85. The value adjusted outcome contains almost four times (3,85) more value in euro’s than the outcome of the best predictive model.

0 10 20 30 40 50 60 70 80 90 100 0 1 2 3 4 5 6 7 8 9 10

Lift Curve

(27)

27

5. Conclusions and discussion

In this chapter, the results will be interpreted and linked to the theoretical assumptions made in the first two chapters. All results that are related to the research questions as well as all results from the benchmark that stand out, will receive the most attention. First, the more general theoretical assumptions tested by means of the benchmark model will be discussed. Second, the social influence variable will be discussed. Third, the value of

longitudinal data will be reviewed. Last, the importance of customer value in churn modelling will be discussed.

5.1 Theoretical assumptions in the benchmark model

Almost all theoretical assumptions made and discussed by means of scientific literature were confirmed by the outcomes of the benchmark model. The few outcomes that not confirm the theoretical assumptions completely will be discussed in this section.

The variable age shows some insignificant age-classes. This insignificance probably relates to the coding of the classes. The general pattern, however, is visible: youngsters churn more and the elder churn less. The same holds for the variable urbanization. Some insignificant classes, but the general pattern is visible: in large cities customers churn more than in small villages.

Another outcome which might be surprising is in the product ranking information. Customers with additional health insurances which are specified at certain target audiences churn more. This is surprising because the product should fulfil the specific needs of the customer who is in this specific target audience. A possible explanation can be found in an article by Ascarza, Iyengar and Schleicher (2016). This article describes that by using price plan

recommendations for your customers (so that they derive greater benefits from the

services), churn goes up. The reason is a lowered inertia/increased awareness, which is also possible due to more personalization of products.

5.2 Research question 1: To what extent does the addition of a (basic) network

variable add value to the predictive power of the churn model (compared to the

benchmark model)?

(28)

28 In order to answer the research question, the predictive power of the model including the social network variable and longitudinal data must be addressed compared to the benchmark model. As can be seen in the results chapter, the predictive power slightly improved (TDL improvement of .05) although this beta is relatively strong.

The reason that the TDL does not improve more than .05 can be found in the number of times the social influence variable contains the value of 1. In the full dataset this number appears only 3458 times: 0,94%. The top decile contains 1261 (36,47%) of the customers who had a co-insured who churned last year, but this is only 3,43% of the total top decile. The overall conclusion of the importance of social network variables is that they are

important (strong, highly significant beta’s), but the way in which the variable is measured in this research does not provide enough observations to strongly influence the predictive power of the model.

5.3 Research question 2: To what extent does historical data add value to the

predictive power of the churn model (compared to the cross-sectional benchmark

model)?

The predictive power of the model is influenced by the addition of historical data, but as already concluded in the previous paragraph: only by a small amount. The reason for this rather small improvement lies in the relatively small number of customers who participated in the changes in collectiveness, voluntary own risk, movements and additional polices in recent history (2012 – 2013) and are still customers in 2014; customers are mostly inert. Some ideas on how to improve the use of historical data will be described in the chapter which deals with limitations and ideas for future research.

5.4 Research question 3: To what extent do the calculated churn probabilities support

retention targeting better when adjusted for customer value?

As can be seen in the results chapter: customer value does matter a lot. The churned

customers in the top decile contain 3,85 times more value than the churned customers in the top decile of the best predictive model. The top decile adjusted for customer value is 91,5% different from the top decile of the best predictive model. Taking customer value into account when making selections for retention campaign can make campaigns more profitable.

5.5 Managerial implications

(29)

29 developments mentioned in this paper. Social influence, historical data and customer value are important for churn predictions and retention targeting.

Although social influence is measured in a limited way, it shows a strong and significant influence on customer churn. It would be valuable for managers to keep in mind that

individual customers do rely on their environment for a certain amount when it comes to the decisions they make. The way the social influence effect can be measured and captured will differ per industry and managers would do good to develop a method their industry/firm fits best.

The data collected on the customer portfolio through the years is valuable for churn

predictions. The beta’s coming from historical data are even stronger than the beta’s coming from recent data, which indicates that a certain amount of customers provides the firm some warning signs before they eventually churn. Especially in markets where customers are less inert it would be wise for managers to check for those warning signs in order to prevent churn in a fast and pro-active way.

The most important finding in this paper is the importance of customer value. Every firm that relies on customers as its primary assets should pay attention to customer value in retention strategies. Campaigns contribute far more to the financial health and success of a firm when accounting for the value customers represent.

5.6 Limitations and ideas for further research

All general phenomena are described and tested with the help of data from a Dutch insurance firm. In order to see how homogeneous the phenomena work over different industries, markets and nations it would be interesting and meaningful to do the same type of research in different settings.

(30)

30 customer value. This approximation assumes that customers who churn on healthcare

insurance, churn on all products. This is short-sighted, but due to circumstances the best way to show a certain way of thought. Another limitation is the static approximation of customer value: this paper uses the fees paid and ignores all opportunities and threats in the development of the customer value over the years. It would be interesting to see what the impact is of adjusting for customer value in a more correct and dynamic way.

(31)

31

Appendix 1: the initial independent variables of the benchmark model

Independent variable Measure Description

Age Categorical Age of the customer Gender Binary Gender of the customer

Social class Categorical The social class of the customer on a household level

Type of residence Categorical The type of residence a customer is living in

Education level Categorical The highest obtained degree of the customer on a household level Relationship length Numeric Length of the customer – firm

relationship in years (on healthcare) Urbanization Categorical Indicates how large the population of

the customers’ hometown is Churned before Binary Indicates whether the customer

churned on his health insurance before Introduction discount Binary Indicates whether the customer made

use of introduction discount (10%) last year

Collectively insured Binary Indicates whether the customer is collectively insured

Became collectively insured Binary Indicates whether the customer became collectively insured last year Voluntary own risk Binary Indicates whether the customer has a

voluntary own risk

Voluntary own risk raised Binary Indicates whether the customer raised his/her voluntary own risk last year Additional health insurances Categorical Indicates which additional health

insurance the customer has (standard, better, specified for an audience, paid bymunicipalities for the minimum incomes)

Dental insurance Categorical Indicates which additional dental insurance the customer has Negative ahi change Binary Indicates whether the customer

dumped his/her additional insurances last year

Moved Binary Indicates whether the customer moved last year

Became main insured Binary Indicates whether the customer became the main insured in 2014 Co-insured adult Binary Indicates whether the main insured has

an adult co-insured

(32)

32

Co-insured ‘younger child’ Binary Indicates whether the main insured has a co-insured between 1 and 11 Co-insured newborn Binary Indicates whether the main insured has

a co-insured between 0 and 1 Car insurance Binary Indicates whether the customer has a

car insurance

Fire insurance Binary Indicates whether the customer has a fire insurance

Number of unique policies Categorical Gold > 4, Silver: 4-3, Bronze: 2, Nickle: 1

(33)

33

Appendix 2: the extra variables added to the total dataset

Independent variable Measure Description

Became collectively insured 2012 Binary Indicates whether the customer became collectively insured in 2012 Became collectively insured 2013 Binary Indicates whether the customer

became collectively insured in 2013 Voluntary own risk raised 2012 Binary Indicates whether the customer raised

his/her voluntary own risk in 2012 Voluntary own risk raised 2013 Binary Indicates whether the customer raised

his/her voluntary own risk in 2013 Negative ahi change 2012 Binary Indicates whether the customer

dumped his/her additional health insurances in 2012

Negative ahi change 2013 Binary Indicates whether the customer dumped his/her additional health insurances in 2013

Moved in 2013 Binary Indicates whether the customer moved in 2013

Social influence Binary Indicates whether a co-insured churned last year

Change in the number of unique policies since last year

Categorical Shows the difference in the number of policies from 2013 and 2014 (can be negative)

Number of times voluntary own risk raised since 2011

(34)

34

Appendix 3 the cleaned benchmark model ~ validation

Model Summary Step

-2 Log likelihood

Cox & Snell R Square

Nagelkerke R Square 1 24921,245a _,131 _,174

a. Estimation terminated at iteration number 5 because parameter estimates changed by less than ,001. Classification Tablea Observed Predicted ind_uitstroom _Percentage Correct nee ja Step 1 ind_uitstroo m nee 6140 3860 61,4 ja 3264 6736 67,4 Overall Percentage _64,4

a. The cut value is ,500 Goodness of Fita

Value df Value/df Deviance 12384,977 10049 1,232 Scaled Deviance _12384,977 ₁₀₀₄₉

Pearson Chi-Square 10371,739 10049 1,032 Scaled Pearson Chi-Square _10371,739 ₁₀₀₄₉

Log Likelihoodb

-8037,530 Akaike's Information

Criterion (AIC) 16143,059 Finite Sample Corrected AIC

(AICC) 16143,178

Bayesian Information

Criterion (BIC) 16411,778 Consistent AIC (CAIC) _16445,778 Dependent Variable: ind_uitstroom

Model: (Intercept), lftcat, ind_tussentweg, ind_kennismaking, ind_coll, ind_er, er_omhoog, av_cat, ind_avweg, ind_verhuisd, SocialeKlasse, Urbanisatie, klantwaarde, ind_auto, ind_meevzVolw, zorgduur

a. Information criteria are in smaller-is-better form.

(35)

35

Appendix 4 The cleaned benchmark model betas

Parameter Estimates

Parameter B Std. Error

95% Wald Confidence Interval Hypothesis Test

Exp(B)

95% Wald Confidence Interval for Exp(B)

(36)

(37)

37 [Urbanisatie=6] ,242 ,0762 ,093 ,392 10,080 1 ,001 1,274 1,097 1,479 [Urbanisatie=5] ,280 ,0588 ,165 ,395 22,695 1 ,000 1,323 1,179 1,485 [Urbanisatie=4] -,021 ,0493 -,117 ,076 ,174 1 ,677 ,980 ,889 1,079 [Urbanisatie=3] ,091 ,0487 -,004 ,187 3,516 1 ,061 1,096 ,996 1,205 [Urbanisatie=2] ,004 ,0516 -,097 ,105 ,006 1 ,939 1,004 ,907 1,111 [Urbanisatie=1] -,065 ,0542 -,171 ,042 1,426 1 ,232 ,937 ,843 1,042 [Urbanisatie=0] 0a _. _. _. _. _. _. ₁ _. _. [klantwaarde=3] -,366 ,0748 -,513 -,220 23,969 1 ,000 ,693 ,599 ,803 [klantwaarde=2] -,250 ,0577 -,363 -,137 18,730 1 ,000 ,779 ,696 ,872 [klantwaarde=1] -,009 ,0454 -,098 ,080 ,039 1 ,843 ,991 ,907 1,083 [klantwaarde=0] 0a _. _. _. _. _. _. ₁ _. _. [ind_auto=1] -,465 ,0541 -,571 -,359 74,049 1 ,000 ,628 ,565 ,698 [ind_auto=0] 0a _. _. _. _. _. _. ₁ _. _. [ind_meevzVolw=1] ,275 ,0428 ,191 ,358 41,105 1 ,000 1,316 1,210 1,431 [ind_meevzVolw=0] 0a _. _. _. _. _. _. ₁ _. _. zorgduur -,539 ,0459 -,629 -,449 137,599 1 ,000 ,583 ,533 ,638 (Scale) 1b

Dependent Variable: ind_uitstroom

(38)

38

(39)

39

Appendix 5 the full benchmark model ~ validation

-2 Log likelihood

a. Estimation terminated at iteration number 5 because parameter estimates changed by less than ,001. Goodness of Fita Value df Value/df Deviance 23783,882 19026 1,250 Scaled Deviance _23783,882 ₁₉₀₂₆ Pearson Chi-Square 19187,986 19026 1,009 Scaled Pearson Chi-Square _19187,986 ₁₉₀₂₆

Log Likelihoodb

Criterion (AIC) 24417,400 Finite Sample Corrected

AIC (AICC) 24417,843 Bayesian Information

Model: (Intercept), lftcat, ind_tussentweg, ind_kennismaking, ind_coll, ind_er, er_omhoog, av_cat, ind_avweg, ind_verhuisd, SocialeKlasse, Urbanisatie, klantwaarde, ind_auto, ind_meevzVolw, zorgduur, ind_collgeworden, tv, ind_polisnemergeworden,

WoningType, Opleiding, ind_brand, ind_meevzKindJong, ind_meevzKindOud, ind_geboorte

a. Information criteria are in smaller-is-better form. b. The full log likelihood function is displayed and used in computing information criteria.

Classification Tablea Observed Predicted ind_uitstroom _Percentage Correct nee ja Step 1 ind_uitstroo m nee 6188 3812 61,9 ja 3283 6717 67,2 Overall Percentage _64,5

(40)

40

Appendix 6: the uncleaned model with social influence and longitudinal data

-2 Log likelihood

a. Estimation terminated at iteration number 5 because parameter estimates changed by less than ,001. Goodness of Fita Value df Value/df Deviance 15138,965 12255 1,235 Scaled Deviance _15138,965 ₁₂₂₅₅ Pearson Chi-Square 12641,621 12255 1,032 Scaled Pearson Chi-Square _12641,621 ₁₂₂₅₅

Log Likelihoodb

(AICC) 18181,728

Model: (Intercept), lftcat, ind_tussentweg, ind_kennismaking, ind_coll, er_omhoog, av_cat, ind_avweg, ind_verhuisd, SocialeKlasse, Urbanisatie, klantwaarde, ind_auto, ind_meevzVolw,

ind_collgeworden12, er_omhoog13, ind_avweg13, ind_avweg12, ind_verhuisd13, zorgduur, klantwaarde_stap_13_naar_14, ind_collgeworden13, ind_er, er_omhoog12, ind_socialinfl, aantal_keer_er_verhoogd_sinds_2011

b. The full log likelihood function is displayed and used in computing information criteria. Classification Tablea Observed Predicted ind_uitstroom _Percentage Correct nee ja

Step 1 ind_uitstroom nee 6337 3663 63,4

ja 3369 6631 66,3

Overall Percentage _64,8

(41)

41

Appendix 7: the cleaned model with social influence and longitudinal data

-2 Log likelihood

a. Estimation terminated at iteration number 5 because parameter estimates changed by less than ,001. Goodness of Fita Value df Value/df Deviance 14575,254 11809 1,234 Scaled Deviance _14575,254 ₁₁₈₀₉ Pearson Chi-Square 12184,141 11809 1,032 Scaled Pearson Chi-Square _12184,141 ₁₁₈₀₉

Log Likelihoodb

(AICC) 17750,462

Model: (Intercept), lftcat, ind_tussentweg, ind_kennismaking, ind_coll, av_cat, ind_avweg, ind_verhuisd, SocialeKlasse, Urbanisatie, klantwaarde, ind_auto, ind_meevzVolw, ind_collgeworden12, ind_avweg13, ind_avweg12, ind_verhuisd13, zorgduur, klantwaarde_stap_13_naar_14, ind_socialinfl,

aantal_keer_er_verhoogd_sinds_2011

b. The full log likelihood function is displayed and used in computing information criteria. Classification Tablea Observed Predicted ind_uitstroom _Percentage Correct nee ja Step 1 ind_uitstroo m nee 6340 3660 63,4 ja 3368 6632 66,3 Overall Percentage _64,9

(42)

42

Appendix 8: parameters cleaned model (Social and Longitudinal)

Parameter Estimates

Parameter B Std. Error

95% Wald Confidence

Interval Hypothesis Test

Exp(B)

95% Wald Confidence Interval for Exp(B)

Lower Upper

Wald

(43)

(44)

44

aantal_keer_er_verhoogd_sinds_2011 ,404 ,0816 ,244 ,564 24,504 1 ,000 1,498 1,276 1,758

(Scale) 1b

Dependent Variable: ind_uitstroom

Model: (Intercept), lftcat, ind_tussentweg, ind_kennismaking, ind_coll, av_cat, ind_avweg, ind_verhuisd, SocialeKlasse, Urbanisatie, klantwaarde, ind_auto, ind_meevzVolw, ind_collgeworden12, ind_avweg13, ind_avweg12, ind_verhuisd13, zorgduur, klantwaarde_stap_13_naar_14, ind_socialinfl, aantal_keer_er_verhoogd_sinds_2011

(45)

45

References

Ascarza, Eva, Raghuram Iyengar, and Martin Schleicher. "The perils of proactive churn prevention using plan recommendations: Evidence from a field experiment." Journal of Marketing Research 53.1 (2016): 46-60.

Ascarza, Eva, and Bruce GS Hardie. "A joint model of usage and churn in contractual settings." Marketing Science 32.4 (2013): 570-590.

Bolton, Ruth N. "A dynamic model of the duration of the customer's relationship with a continuous service provider: The role of satisfaction." Marketing science 17.1 (1998): 45-65. Braun, Michael, and David A. Schweidel. "Modeling customer lifetimes with multiple causes of churn." Marketing Science 30.5 (2011): 881-902.

Ferguson, Rick, and Bill Brohaugh. "Telecom's search for the ultimate customer loyalty platform." Journal of Consumer Marketing 25.5 (2008): 314-318.

Gupta, Sunil, Donald R. Lehmann, and Jennifer Ames Stuart. "Valuing customers." Journal of marketing research 41.1 (2004): 7-18.

Haenlein, Michael. "Social interactions in customer churn decisions: The impact of

relationship directionality." International Journal of Research in Marketing 30.3 (2013): 236-248.

Holtrop, Niels, et al. "No future without the past? Predicting churn in the face of customer privacy." International Journal of Research in Marketing (2016).

Knox, George, and Rutger Van Oest. "I'm Here, But Maybe Not for Long: Using Complaints to Predict and Value Customers." Available at SSRN 1032843 (2007).

Leeflang, P.S.H., Wieringa, J.E., Bijmolt, T.H.A. & Pauwels, K.H. (2015) Modeling Markets Analyzing Marketing Phenomena and Improving Marketing Decision Making, 1st edition, New York: Springer.

Lemmens, Aurélie, and Christophe Croux. "Bagging and boosting classification trees to predict churn." Journal of Marketing Research 43.2 (2006): 276-286.

(46)

46

Lemon, Katherine N., Tiffany Barnett White, and Russell S. Winer. "Dynamic customer relationship management: Incorporating future considerations into the service retention decision." Journal of marketing 66.1 (2002): 1-14.

Mittal, Vikas, and Wagner A. Kamakura. "Satisfaction, repurchase intent, and repurchase behavior: Investigating the moderating effect of customer characteristics." Journal of marketing research 38.1 (2001): 131-142.

Neslin, Scott A., et al. "Defection detection: Measuring and understanding the predictive accuracy of customer churn models." Journal of marketing research 43.2 (2006): 204-211. Nitzan, Irit, and Barak Libai. "Social effects on customer retention." Journal of Marketing 75.6 (2011): 24-38.

Risselada, Hans, Peter C. Verhoef, and Tammo HA Bijmolt. "Staying power of churn prediction models." Journal of Interactive Marketing 24.3 (2010): 198-208.

Romp, Marnix and Merkx, Paul (2015), “Zorgthermometer,” (accessed September 19, 2016), [available at http://www.vektis.nl].

Schweidel, David A., Peter S. Fader, and Eric T. Bradlow. "Understanding service retention within and across cohorts using limited information." Journal of Marketing 72.1 (2008): 82-94.

Srinivasan, Shuba, and Dominique M. Hanssens. "Marketing and firm value: Metrics, methods, findings, and future directions." Journal of Marketing research 46.3 (2009): 293-312.

Verbeke, Wouter, et al. "New insights into churn prediction in the telecommunication sector: A profit driven data mining approach." European Journal of Operational Research 218.1 (2012): 211-229.

Verhoef, Peter C., Philip Hans Franses, and Janny C. Hoekstra. "The effect of relational constructs on customer referrals and number of services purchased from a multiservice provider: does age of relationship matter?." Journal of the Academy of Marketing Science

30.3 (2002): 202-216.