Master’s Thesis MSc Marketing 2020
Will you stay or will you go?
Predicting churn in a Dutch secondary schoolbook marketAndries Elsinga S3793524 Supervisor: Abhi Bhattacharya
Faculty of Economics and Business
2
Will you stay or will you go?
Predicting churn in a Dutch secondary schoolbook marketMaster thesis MSc Marketing intelligence 22 May 2020 Andries Elsinga S3793524 Morielje 11 9404 MD Assen a.elsinga.1@student.rug.nl +31640257467 University of Groningen Faculty of Economics & Business
Department of Marketing PO Box 800, 9700 AV Groningen (NL)
Supervisors:
First: Dr. Abhi Bhattacharya
abhi.bhattacharya@rug.nl
Second: Prof. Dr. Jaap Wieringa
j.e.wieringa@rug.nl
Noordhoff: Dr. Sander Beckers
3
Summary
This paper will research which factors serve as the best predictors when it comes to churn in the Dutch secondary schoolbook market. This research will be carried out in a schoolbook publishing company providing schoolbooks and digital learning materials to elementary schools, secondary schools, higher education and the professional market in the Netherlands. The goal of this paper is to examine the effect of customer service contact and previous churn behavior on churn. The research question of this paper is: “What is the effect of customer service contact and previous churn behavior on churn probabilities?”
A predicative multilevel model is used to estimate which factors are good predictors for churn. This model will provide estimates of the effects of customer service contact on churn and their interaction effect with previous churn behavior and nature of contacts. A multilevel model will be used to predict churn, since the data used for analysis is nested and group membership matters in this context.
The results of this study show that recency and previous churn behavior are predictors of churn probability. The more time that has passed between the last contact of a customer with the service desk, the more the churn probability will rise. This indicates that the effect of being in contact with a company and being tied to it by staying in touch fades over time (knox & van Oest, 2014; Umashankar, Ward & Dahl, 2017). Additionally, the more a customer has churned in the past, the lower the churn probability will be. This indicates that customers who have churned a lot know what to expect and settle down for product when it meets their expectations. In this paper we also found that years in use has an effect on the churn
probability. The longer a customer uses a product, the less likely they are to churn. This could be explained by inertia, when long time users are afraid of switching due to the efforts and changes that are inherent to it.
Based on the results of this paper it is recommended to allocate more resource to people who haven’t had recent contact with the company because they are more likely to churn.
Allocating more resources to these customers can tie them to the company. It is also
4
5
Contents
Summary ... 3 Introduction ... 6 Theoretical background ... 10 Churn ... 10Customer service contact ... 10
Nature of contact ... 11
Previous churn behavior ... 12
Model ... 14 Research design ... 15 Data ... 15 Variables ... 16 Descriptive statistics ... 18 Method ... 21 Model specification ... 22 Results ... 25
Outliers and missing data ... 25
Multicollinearity ... 26
Discussion ... 33
References ... 37
6
Introduction
Nowadays, churn prediction is a very relevant topic within customer management (Kumar & Reinartz, 2012). Retaining customers can decrease the need to spend lots of money acquiring new customers in order to keep a positive cash flow. Retaining customers is important because it can lead to considerable financial gains. Businesses typically spend a much larger amount of money than individual consumers making customer retention a key factor in maintaining revenues and business strategies in general (Jahromi, Stakhovych & Ewing, 2014; Boles et al., 1997; Rauyruen & Miller, 2007). In addition, it is financially more beneficial to retain the existing customer base than to spend money trying to acquire new customers (Gupta, Lehmann, & Stuart, 2004; Athanassopoulos, 2000). Predicting customers’ churn probability allows companies to know how to spend their budget more effectively to prevent churn. This paper will examine how to predict whether customers will switch to other suppliers as accurately as possible. This research will be carried out in a schoolbook
publishing company providing schoolbooks and digital learning materials to elementary schools, secondary schools, higher education and the professional market in the Netherlands. The main focus of this paper will be the secondary schools, since this is the most interesting market to the company. This is the most interesting market, because the customers in this market churn most often. Students also don’t have to decide whether to buy the books, since the schools in this market purchase these for them.
The two main sets of data used to predict churn are customer service data and data about the books used by the secondary schools. Customer service data includes detailed information about interactions with customers through the service department. Data about the books used by the secondary schools includes information about which learning methods schools have been using for the past years, thus telling us when and whether they churn. The main effects, that are investigated in this paper, are frequency, recency and nature of customer service contact and previous churn behavior. The effect of nature of customer service contact on churn is expected to be moderated by previous churn behavior and the frequency of customer service contact. The goal of this paper is to examine the effect of customer service contact and previous churn behavior on churn. The research question of this paper is: “What is the effect of customer service contact and previous churn behavior on churn probabilities?”
7
These papers usually explore the reason customers churn and examine ways to either prevent this or win these customers back (Jahromi, Stakhovych & Ewing, 2014; knox & van Oest, 2014; Kumar, Leszkiewicz & Herbst, 2018). Satisfaction and complaints in relation to churn are examples of study’s that have been conducted in B2B markets to explore the reasons for churning and the methods to prevent this (van Doorn & Verhoef, 2008; knox & van Oest, 2014). In both B2C and B2B markets churn is researched in contractual as well as non-contractual settings. Popular examples of markets in which papers study churn in non-contractual settings are telecom markets (Kumar, Leszkiewicz & Herbst, 2018) and banking (Cheng, Wu & Chen, 2019). A key difference between these settings is that churn, in contractual settings, is immediately detected when a contract is cancelled or isn’t renewed. In non-contractual settings it is much more complicated to determine whether a customer has churned. This paper studies the effect of customer service contacts and previous churn behavior on churn in a contractual B2B setting.
Frequency, recency and nature of customer service contacts are researched in this paper along with previous churn behavior. One of the insights this paper will deliver are expected to be knowledge about whether and, if so, how much the frequency of contact with customer services will influence churn probability. Another expected insight is knowing if the recency of the last contact with customer services matters in predicting churn and how large this effect is. Additionally, the effect of nature of contacts on churn will be an insight provided by this study. This insight will tell us whether negative or positive contacts matter when predicting churn. Finally, another insight we could get from this study is the effect of previous churn behavior on churn probability. It will tell us if previous churn behavior matters in a churn prediction and how it affects this prediction.
All these insights will be gathered by analyzing data from the customer service department and data from booklists of secondary schools. While the effect of customer service on churn has been studied before, taking nature of the contact into account can add new insights. Including the nature of the contact in the study could provide interesting knowledge of the possible effects service contacts can have, depending on its nature. The nature of the contact could determine whether the relation between service contact and churn is negative or positive (Retana, Forman & Wu, 2016; knox & van Oest, 2014; van Doorn & Verhoef, 2008).
8
is available at most firms, to predict churn probabilities. The metrics frequency and recency in particular, are very easy to measure in most companies. Measuring nature of contacts can be more challenging depending on the data that is gathered, but it might be useful.
These are important factors to study because, firstly, one can determine whether the nature of customer contacts actually matters when it comes to churn. Several studies have determined effects complaints might have. This paper will provide insights in whether it will actually matter if a contact is negative or positive and, if so, how much this matters. This knowledge is important to allow managers to identify if customers with negative contacts will for example have a higher churn probability, where customers with positive contacts will have lower probabilities. Managers could use this knowledge to more effectively allocate resources to customers with the highest churn probability. It is important to study the effect of recency and frequency to determine how useful these metrics are in predicting customers churn
probabilities. These two metrics are often easily accessible for firms and can be valuable when applying the knowledge gained in this study.
The results of this study show that recency and previous churn behavior are predictors of churn probability. The more time that has passed between the last contact of a customer with the service desk, the more the churn probability will rise. This indicates that the effect of being in contact with a company and being tied to it by staying in touch fades over time (knox & van Oest, 2014; Umashankar, Ward & Dahl, 2017). Additionally, the more a customer has churned in the past, the lower the churn probability will be. This indicates that customers who have churned a lot know what to expect and settle down for product when it meets their expectations. Among the control variables, there was also an interesting effect of years in use on the churn probability. The longer a customer uses a product, the less likely they are to churn. This could be explained by inertia, when long time users are afraid of switching due to the efforts and changes that are inherent to it.
9
10
Theoretical background
Churn
Churn is an interesting concept in a contractual setting. Churn is the event where a customer, who is currently buying from a company, stops buying from this company (Huang, Kechadi, Buckley, Kiernan, Keogh & Rashid, 2010). It is interesting in this setting because churn is immediately registered if a contract is either terminated or not renewed. In non-contractual settings churn can’t be registered that easy. If a customer churns in a contractual setting, it means a customer won’t be back for a while since the customer has a contract somewhere else.
Customer service contact
Customer service contact is defined in this paper as “a customer contacting the customer service desk”. A customer might contact the customer service desk for varying reasons. The most obvious reason to contact customers service is for complaints or technical support. But contacts with customer service can also mean requesting information on the usage of
products, or ordering new products. Current literature offers various perspectives on the relationship between customer service and churn. For example, the study by Slack & Singh (2020) found that customer satisfaction and perceived service quality affect customer loyalty. Lower satisfaction, which is often accompanied by complaints, increase the chance of
churning. In addition, the efforts customer put into complaints can reduce their satisfaction and doing so, increase the chance the customer might churn (Cai & Chi, 2018).
However, other studies imply that contact with the customer service desk actually decrease the chance of churn. Customers who file a complaint even tend to churn less whenever the complaint was addressed often and fast by the service desk (van Doorn & Verhoef, 2008). Also, staying in touch with the company by complaining increases customer loyalty
11
In this paper, customer service contact is measured in two ways. The first way it’s measured is frequency of customer service contact. Frequency of customer service contact is expected to be positively related to churn probability. The more contacts there are between a school and the company the more likely they would be to churn.
Based on literature some could propose that frequency has a negative relation with churn probabilities. Contacting customer services can increase involvement with a company and tie customers to it. Handling these contacts well might improve the relationship with the
customer, further decreasing churn (van Doorn & Verhoef, 2008). However, since we assume the majority of contacts with customer service are initiated to complain or report technical issues, a negative relation is expected. This expectation is based on intuition and the literature that proposes that contacts like complaints can make customers less satisfied with a company and increase their churn probability. The first hypothesis will be formulated as follows: H1: Average frequency of contact with customer service is positively related to churn probability of sections.
The second way in which customer service contact is measured is recency. The recency of a contact with the customer service desk is expected to be positively related to churn. The more time has passed since the last contact with the service desk, the higher the churn probability is. This expectation is based on the theory about customer involvement and being tied to a company by staying in contact with the company. Presumably, if the time that has passed since last contact with a customer is will increase, the less this customer might still be tied to the company and the higher their churn probability will get. Also, the effects of contact with the customer fades over time (knox & van Oest, 2014), implying that the churn reducing effect of contacts proposed by van Doorn & Verhoef (2008) and Umashanka, Ward & Dahl (2017) will also fade away. The hypothesis will be formulated as follows:
H2 : Average recency of contact with customer service is positively related to churn probability of sections.
Nature of contact
12
expression of dissatisfaction. Negative contacts generally increase the chance of churn significantly. This is, however, influenced by previous satisfaction (knox & van Oest, 2014; van Doorn & Verhoef, 2008).
On the other hand, contacting customer services to, for example, obtain more information about usage of a product could have a different nature. A customer might be satisfied with the product and wants to learn how to use the product more or better. These kinds of contacts have been found to be able to increase usage of educational products up to 50 percent. . Involvement plays a part in this effect and his kind of involvement leads to a lower churn probability (Retana, Forman & Wu, 2016).
Because we expect the effect of nature of contacts on churn probabilities to depend on the type of nature, the expectation is that nature of contacts is negatively related to churn probability. When contacts of a school are positive, their churn probability is lower. If contacts are negative, churn probabilities increase. The hypothesis will be formulated as follows:
H3: Average nature of contact with customer service is negatively related to the churn probability of sections.
We expect negative contacts to increase churn probability and we expect positive contacts to decrease this probability. It seems a logical assumption that, if the amount of contacts and the nature of contacts have an effect on churn, the amount of contacts of a specific nature have an effect as well. If a customer has mostly negative contacts with the company, and this customer often has these issues, it seems likely that the churn probability of this customer would be higher than the churn probability of a customer that rarely contacts the company with these kind of issues. Therefore, we expect the frequency of customers service contacts to strengthen the relationship between nature of contacts and churn probability. The following hypothesis is formulated:
H4: Average frequency of customer service contact positively influences the effect of average nature of contact with customer service on the churn probability of sections.
Previous churn behavior
13
churn often tend to churn repeatedly (Kumar, Leszkiewicz & Herbst, 2018). Switching costs are costs one makes to switch from the products of one supplier to another. Switching costs decrease the probability of churning. High switching costs can lead to customers buying from a certain company even when there are better alternatives (Cullen and Shcherbakov 2010). Customers who switch often tend to experience these switching costs as lower than the
customers who switch rarely (Bell, Auh, and Smalley 2005; Yanamandaram and White 2006). The regular churners know the process they need to go through to switch to another product. Customers who rarely churn are less familiar with this process are more held back when it comes to churning. Because these regular churners are familiar to the switching process, they are likely to have a higher churn probability. That’s why previous churn behavior is expected to be positively related to churn probability. Churning frequently increases the probability of churning. The hypothesis is formulated as follows:
H5: Previous churn behavior of sections is positively related to the churn probability of sections.
Previous churn behavior is also expected to have an effect on the relation between nature and churn. Since repeated churners are less held back to churning, one could argue that these customers will have a higher churn probability than customers who rarely churn when they have more negative contacts. Whenever a customer has mostly negative contacts, the effect of these contacts will be stronger whenever the customer has churn more often in the past. The effect of previous churn behavior is therefore expected to strengthen the relation between nature and churn probability. The hypothesis is formulated as:
14
Model
In figure 1, the conceptual model is displayed. For every variable, the level is stated in between brackets. The variables are either group variables (school level) or individual
variables (section level). Also, for every variable the expected relationship is stated by adding a plus or minus sign in between brackets. Frequency and recency of customer service contact are expected to have a positive relation with churn probability. Nature of customer service contact is expected to have a negative relation with churn probability. Previous churn behavior is expected to increase churn probability.
Frequency of customer service contact (school level) ( + )
Churn probability (section level)
Control variables Previous churn behavior
(section level) ( + )
Figure 1: Conceptual model
H2
H5 Nature of contact (school
level) ( - )
H1 H3 H6
Recency of customer service contact (school level) ( + )
15
Research design
Data
For this paper, data from 22.286 sections is used for analysis. The data from secondary schools is used for analysis, because this is the largest segment the company serves. In addition, these schools can renew methods more often than primary schools (about every 4 years, primary schools ever 8-9 years). Secondary schools purchase books for every student, where in professional and higher education, students need to buy books themselves. They can choose not to buy a book, or buy it second hand. Secondary schools are the most interesting for the company and churn effects can be investigated the best in this market due to the higher frequency of churn and the schools deciding whether to buy the books. The products are sold to the schools on contractual basis.
The data is available on different levels. The first level is section level. Which means the data is aggregated per course, stream, study phase and school. The second level is per school. Customer service data is available on school level and churn data is available on section level. We can’t assume that sections within schools are as independent from each other as they are from sections from other schools. It is likely that the section within a school are more similar to each other than to sections of other schools.
The churn data is obtained through collecting book list data from schools. This tells the company which books every school uses in every school year. Since the books are sold in a contractual setting, churn is detected when the book for a section is different from the book that section put on the book list in the previous year. Churn and previous churn behavior can be measured like this.
The data of all available schools was included in the research. Afterwards, observations with missing data on churn variables were excluded. This results in a group of 22.286 sections being analyzed in this study. All sections are customers of the company at the time of this research. The study sample consists of nearly every section that is currently using methods from the company. This makes sense, since the goal of the study is to predict churn for all current customers. Including almost every current customer makes the sample representative since it is almost the entire population that is studied.
16
with the customer. Whenever a customer calls, emails or writes about something, it is
registered as a ticket. In this ticket all data about that subject is stored, given that it is properly logged by the customer service agent. The date a ticket was created and solved is registered along with the subject of the ticket. Also, the school who the person contacting the
ServiceDesk is working for is registered. Using this data, recency and frequency of tickets per school can be measured.
To measure the nature of the tickets, the subjects of the tickets are examined. These subjects are classified as negative, positive or neutral. Using the Kappa statistic (Cohen’s Kappa) for inter rater reliability, the validity of these classifications is measured. The Kappa statistic has a value between -1 an 1 where -1 is a perfect disagreement and 1 a perfect agreement
(MChugh, 2012). When the statistic is between 0,61 and 0,80 a substantial agreement is reached, where a statistic above 0,80 can be considered near perfect (Landis & Koch, 1977). Kappa statistics below 0,80 can often still be accepted within marketing studies. A guideline by Fleiss (1981) indicates 0,75 to be an acceptable cutoff point. The Kappa statistic is
normally used for 2 raters. Rust and Cooil (1994) describe how this statistic can be calculated for multiple raters.
The Kappa statistic was calculated by analyzing the ratings of three raters. All raters classified 51 ticket subjects as negative, neutral or positive. To test whether there was a significant difference in rating between these raters a ANOVA was conducted. This test showed that there was no significant difference in the rating given by the three raters to the subjects, with a P-value of 0.8649.
Using the method by Rust and Cooil (1994), the number of agreements between the three raters is counted for each subject. The sum of all agreements is divided by the maximum possible amount of agreements in the ratings. This yields the proportion of interjudge
agreement. In this research the proportion of interjudge agreement is 0.93. Using this number, a PRL reliability can be extracted from a table provided in the article which, in this case, is 100. This indicates that there is a very high inter rater reliability.
Not all subjects were classified in the same way by all three raters. For the cases where one rater rated the subject differently, the subject was classified by following the rating of the two raters that did agree.
Variables
17
Variable Measurement
Dependent variable
Churn Binary variable measuring whether the section has churned in school year 2018/2019.
Main effects
Recency Continuous variable measuring the amount of days that have passed since the last ticket was opened until the end of the school year (31/07/2019).
Frequency Continuous variable measuring the amount of tickets opened per school from the end of the 2016/2017 school year.
Nature Ratio variable measuring the nature of tickets per school since the 2016/2017 school year.
Moderators
Previous churn behavior Continuous variables measuring the amount of times a section has churn in the last 6 school years.
Control variables
Years in use Continuous variable measuring the amount of sequential years that a section has used their current method.
School size Continuous variable measuring the amount of students per school in the school year 2018/2019.
Cross buying Ratio variable ranging from 0 to 1 measuring the ratio of the companies methods used by a school out of al methods the school uses.
Newsletters Continuous variable measuring the amount of newsletter received via email per school since the end of the 2016/2017 school year.
18
Frequency online usage Continuous variable measuring the amount of scores set in an online learning method in the school year 2018/2019.
Dips and peaks Continuous variable measuring the difference between the monthly frequency of online learning scores peaks and dips. Where peaks are measured by the monthly frequency of online learning scores being at least 1 standard deviation below the annual mean and dips are measured by being below the first quartile of the annual mean. These measurements are for the school year 2018/2019.
Course Nominal variable indicating the subject that is taught by the section.
Table 1: variable measurements.
Descriptive statistics
Before analyzing the data, some descriptive statistics are examined to check the data. First, churn rates per subject and per education type are displayed and reviewed. Next, some
descriptives regarding the main effects, moderators and churn are explored. Lastly, cross level relationships in the variables are reviewed to determine the relevance of a multilevel model. In table 2 the churn rates per subject is summarized. The 11 subjects that have the highest amounts of observations (more than 500) were included in the table. These subject are generally taught in almost every school because they are a mandatory part of most
curriculums. The subjects that are not included are usually optional or very specific subjects for certain stream on VMBO. The table shows that there is some variation in churn among the subjects, ranging from a low churn rate of 2.22% to a moderately high churn rate of 8.12%. Despite these differences, most of the churn rates are relatively close to the total average churn rate of 4.74%, telling us that churn is somewhat constant among the sections. The table also shows that even among the subjects with more than 500 observations, there is quite a difference in amount of sections teaching the subject.
Subject Amount of sections teaching this subject
Churned sections Churn rate
Total 21783 1033 4.74%
Geology 1160 44 3.79%
19 German 2672 156 5.84% Economics 1669 37 2.22% English 1871 117 6.25% French 1616 43 2.66% History 1108 90 8.12% Physics 745 49 6.58% Dutch 2518 87 3.46% Chemistry 1285 86 6.69% Mathematics 4198 131 3.12%
Table 2: churn rate per course.
In table 3 below, an overview of churn rates per education type is displayed. The education types are the three streams in the Dutch education system: VMBO, HAVO and VWO combined with the study phase. The study phases are OB, onderbouw in Dutch, junior in English, and BB, bovenbouw in Dutch and senior in English. The table shows that VMBO OB and BB are the largest streams and that HAVO and VWO are quite similar in size. The churn rates among the education types don’t differ so much from each other, but they are a bit lower for VMBO.
Section Amount of sections Churned sections Churn rate
VMBO OB 6132 277 4.52% VMBO BB 5450 237 4.35% HAVO OB 3063 171 5.58% HAVO BB 2524 128 5.07% VWO OB 2742 147 5.36% VWO BB 2415 120 4.97%
Table 3: Churn rate per education type.
In table 4 some descriptive statistics of the main effects, moderators and the dependent
20
that often. Looking at previous churn rates we see a minimum of 0 and a maximum of 5, which makes sense since it is the churn calculated over the last 6 years. The mean is 1.25 and the standard deviation is 0.606. These measures also make sense, because most schools can only afford to churn after about 4 years due to budget limitations. The median is 1, indicating that more than half of the sections churn at least once in these 6 years. Looking at the range and mean of recency, we see that the most time passed since the last ticket can be high (about 3 years) and the least time is one day. The average amount of time passed since the last ticket is 43 days, which seems reasonable. The median is 23, which tells us that most observations are within a month. Frequency ranges from 1 to 251 with the average being 85.221. Looking at the standard deviation of 46.720 we see that frequency varies quite a bit. A possible reason might be that some schools are larger or have been a customer for a long time and therefore accumulate more tickets. Looking at nature of tickets, the table shows that the minimum is -1, meaning that a school has only negative tickets and a maximum of 0,5, meaning they have a large proportion of positive tickets. The mean of nature is negative. This indicates that on average, schools have more negative than positive tickets. An explanation could be that customers usually contact customers service when something is wrong and not so much about positive subjects (Alsem & Klein Koerkamp, 2016).
Variable Min Max Standard
deviation Mean Median Churn 0 1 0.213 0.048 0 Previous churn behavior 0 5 0.606 1.25 1 Recency 1 1067 80.544 43.387 23 Frequency 1 251 46,720 85.221 82 Nature -1 0.5 0.176 -0.103 -0.093
Table 4: Descriptive statistics. Correlation
21
nature churnbehavior recency frequency
nature 1.00 -0.05 * -0.03 * 0.03 *
churnbehavior -0.05 * 1.00 -0.01 0.04 *
recency -0.03 * -0.01 1.00 -0.36 *
frequency 0.03 * 0.04 * -0.36 * 1.00
Table 5: Correlation matrix.
Method
The goal of this research is to determine the effect of customer service contact on individual churn probabilities of sections and what factors influence this effect. To research this, a predictive model will be estimated. Using the previously described data, this model will provide estimates of the effects of customer service contact on churn and their interaction effect with previous churn behavior and nature of contacts. A multilevel model will be used to predict churn. This type of model will be more accurate in predicting churn than an ordinary linear model (Gelman, 2006). This model will be estimated using logistic regression (LOGIT) in R.
The predictive model is a nested model. The individual observations are section level observations nested into groups, which are schools. In these groups, frequency, recency and nature of contacts are group variables. Because these variables are constant within groups, variance in these variables can’t influence individual differences in the variable parameters within the groups themselves. These variables do affect the individual churn probability within the group, making the individual churn probability to be dependent on group variables. This means group membership matters in this context because these variables can influence differences between groups, meaning a multilevel model is the most appropriate method to estimate churn probabilities.
22
variables but they are not applicable for multilevel models (Leeflang, Wieringa, Bijmolt & Pauwels, 2015). In binary choice models, logistic regression has often been used as a method for predicting churn (Leeflang, Wieringa, Bijmolt & Pauwels, 2015; Risselada et al., 2010). Logistic regression is the most appropriate way to estimate churn, because churn is measured binary, other machine learning methods are not applicable and logistic regression is regarded as a good prediction model for churn.
Model specification
The model is specified in the following equations as a multilevel model. Interaction effects are added to estimate moderation effects. Not only the main effects and moderation effects are included, but also control variables are added. The model is specified on two levels. The first level is section level and the second level is school level. Churn, previous churn behavior, years in use, online usage recency, online usage frequency, online dips and peaks and course are the variables included in level 1. These variables are included in this level because there are all measured at section level, which is nested in school level. The variables included in level 2 are frequency, recency, nature, school size, cross buy and newsletters. These variables were included in level 2 because all of them are measure on a school level.
23 Main effects model
24 Where:
ε = Residual for section i in school j. u = Residual for school j.
Level 1:
C = Churn for section i in school j.
P = Previous churn behavior for section i in school j. Y = Years in use for section i in school j.
OR = Online usage recency for section i in school j. OF = Online usage frequency for section i in school j. ODP = Online usage dips and peaks for section i in school j. V = Course for section i in school j.
Level 2:
25
Results
Outliers and missing data
Before the data was analyzed, outliers and missing data were dealt with. These were either removed or transformed. This section will explain how outliers and missing data were dealt with.
Frequency
Frequency is the measure of the amount of tickets that are linked to one school. While looking at the data, it was obvious that there were a few values that were sticking out. As shown is figure 2, the left boxplot displays some values that are much higher than most other values. Any frequency value above 10.000 was removed from the data since it seems unreasonable that one school has accumulated so many tickets. After removing these 7 observations, the boxplot looks a lot better (right on figure 2).
Figure 2: Boxplot of frequency before and after removal of outliers. Previous churn
26
influences the previous churn behavior. If one or more years are missing for a section, it is automatically measured as churn if we measure churn by looking if a section uses a different method then the previous year. Another method would be to only count the churning in the years where we have data. This would give us a more reliable value. However if we measure this over the last 12 years, the values would become more unreliable since as we go further back in time, more values are missing. Therefore we only count the churn behavior in the past 6 years. This would measure if a section churns often (use method for less than 4 year), if a section churns as expected (uses method for 4 years) or if a section churn rarely (doesn’t churn after 6 years). This is a reasonable timeframe where we have very little missing data compared to the data from the past 12 years.
Learning data
Learning data was included as a control variable. The data was gathered and transformed in a per school, per method per month level. It was afterwards aggregated to section level to be used for analysis. Especially the frequency of scores in online learning modules had quite some outliers. The distribution of frequency was non-normal. Most observation where low frequencies but there were some frequencies that were very high. To make sure dips and peaks in usage were properly registered, some data was removed and variables were changed. Firstly, all frequency observations above 415 scores were removed, which was about 5% of the observations. Since the observations above 415 were generally very high with a maximum of 559.263 scores, these observations influenced the standard deviation very much. After removing these outliers, the standard deviation was still larger than the mean and thus, was not fit to use for computing the dips variable (one standard deviation below the mean). Therefore, the dip variable was redefined as all observations below the first quartile of the data. Peaks were still measured as observation with one standard deviation above the mean.
Multicollinearity
27
To solve the problem of multicollinearity, the nature of ticket variables were combined in one single variable. This variable is computed in a similar way a net promotor score is calculated. First, however, positive, negative and neutral tickets were divided by the sum of all 3
categories. The ratio of negatively classified tickets were then subtracted from the ratio of positive ones. By doing this, the resulting variable will be less likely to follow a similar pattern to frequency. Variable VIF Frequency 170.295 Recency 1.527 Neutral tickets 31.040 Positive tickets 17.661 Negative tickets 28.288
Previous churn behavior 1.065
Years in use 1.221 School size 1.220 Newsletters 2.821 Crossbuy 1.254 Recency LD 1.172 Frequency LD 1.566 Peak 1486662 Dip 1.434 Course 3.250
Table 6: VIF values.
In addition to the multicollinearity issues with nature of tickets, we can see that peaks also suffer from multicollinearity. Their VIF value is much higher than the threshold, so they need to be adapted. Peaks and dips are combined into one variable to prevent them to cause issues in the estimation of the model. The new variable is the amount of peaks subtracted by the amount of dips. After replacing peaks and dips with this variable, the model doesn’t suffer from multicollinearity anymore, as shown in table 7. All VIF values are below 4.
Variable VIF
Frequency 2.694
28
Nature 1.234
Previous churn behavior 1.053
Years in use 1.204 School size 1.094 Newsletters 2.598 Crossbuy 1.197 Recency LD 1.172 Frequency LD 1.555 Peak.dip 1.419 Course 2.967
Table 7: VIF values
Estimating the model in R
To avoid scaling issues in R, the indepdendent variables, expect for courses, where z-scored. To estimate the model in R, the lme4 package was used. First, a generalized linear mixed effects model was estimated using the glmer function as a LOGIT. Previous churnbehavior was used as a random slope in the model and recency, frequency and nature were used as random intercepts. One model was estimated for the main effects and one model was
estimated for the interaction effects. Several optimizers where estimated to determine which LOGIT model had the best fit. For both models, three optimers where compared. A
comparison of the models is shown in tables 8 and 9. The main effects model using the “optimx” optimizer preforms the best, looking at the BIC. This model will be used for
estimation. For the moderator model “optimx” and “bobyqa” perform equally well. However, “optimx” yields no warnings in R, where “bobyqa” does. Therfore “optimx” will be used for estimation.
Main effects model AIC BIC LL
Optimx 368.1 523.3 -156.0
Bobyqa 368.1 523.3 -156.0
Nelder_mead 368.9 524.0 -156.4
nlopt 368.2 523.4 -156.1
Table 8: Comparison between optimizers for the main effects model.
29
Optimx 369.9 547.3 -153.0
Bobyqa 369.9 547.3 -153.0
Nelder_mead 371.2 548.5 -153.6
nlopt 370.1 547.4 -153.0
Table 9: Comparison between optimizers for the moderator model.
In table 10, the results of the main effects model are shown and in table 11 the results of the moderator effects are shown. There are some significant parameters. We can see that previous churn behavior has a marginally significant negative effect on churn, recency has a marginally significant positive effect on churn and years in use has a significant negative effect on churn. Looking at the interaction effects, we see that none of them are significant. This indicates that there are no moderator effects that can be confirmed.
Parameter Odds ratio Estimate Standard error T value P value
(Intercept) 0.0003 -8.166 1.197 -6.823 0.000 *
Nature 1.039 0.038 0.449 0.084 0.933
Previous churn behavior 0.287 -1.248 0.678 -1.840 0.066 *
Recency 2.058 0.722 0.425 1.697 0.089 * Frequency 0.956 -0.045 0.635 -0.070 0.944 Years in use 0.218 -1.525 0.433 -3.519 0.000 * School Size 1.276 0.243 0.384 0.634 0.526 Newsletters 2.150 0.766 0.518 1.478 0.140 Crossbuy 0.613 -0.489 0.412 -1.186 0.236
Recency Online Learning 0.775 -0.255 0.425 -0.599 0.549
Peaks and dips Online
Learning 0.990 -0.010 0.280 -0.035 0.972
Frequency Online Learning 0.673 -0.397 0.375 -1.058 0.290
30 Course: Social studies 878135000
000 27.501 14167.08 0.002 0.998
Course: Human and social
studies 0.000 -15.546 25228.51 -0.001 0.999
Course: Physics & chemistry 116.231 4.756 1.791 2.655 0.008 *
Course: Physics 0.000 -16.716 23750.94 -0.001 0.999
Course: Dutch 1.905 0.644 1.016 0.634 0.526
Course: Chemistry 9.681 2.270 1.070 2.121 0.034 *
Course: Mathematics 2.313 0.838 1.159 0.723 0.469
Table 10: Main effects model (optimx) without robust standard errors.
Parameter Odds ratio Estimate Standard error T value P value Previous churn behavior
:Frequency 1.140 0.131 0.768 0.170 0.865
Frequency:Nature 0.422 -0.863 0.475 -1.816 0.069 *
Table 11: Moderator model (optimx) without robust standard errors. Hypothesis testing
Now that the results are clear, the hypothesis can be tested. In table 12 it is shown whether the hypothesis is confirmed or not. Of all hypotheses that were formulated, two could be
confirmed in this research. One main effect and a moderator showed a significant relation with churn. The interaction effects are insignificant and the moderators have no significant relations with the main effects.
It is interesting to see that, not like expected, there is no significant relation between
31
There is no significant effect of nature on churn. That means that H3 is not confirmed. What is interesting to see is that there is a marginally significant interaction effect between
frequency and churn. However, because there is neither a significant effect of nature or frequency on churn, H4 can’t be confirmed.
We can see a marginally significant negative relation of previous churn behavior with churn with a P value of 0.066. This indicates the longer a section uses a certain method, the lower the chance of this section churning. The expectation in H5 was that there would be a relation between these 2 variables, but this relation was expected to be positive. Since the relation is marginally significant, the opposite of H5 can be confirmed.
Among the control variables there are also some significant parameters. These parameters are not related to the hypotheses, but they still have some interesting results. We see a significant negative effect of years in use on churn with a P value smaller than 0.001. The relation of years in use and churn is interesting because it is a negative effect. This indicates that the longer a section uses a certain learning method, the less chance there is of churning. There are also some significant effects of certain courses. The significant parameters for course all have positive coefficients. This indicates that English, Physics & chemistry and Chemistry are all course where sections generally churn more often.
Hypothesis Relation Confirmed
H1 Frequency (+) churn No
H2 Recency (+) churn Yes
H3 Nature (-) churn No
H4 Frequency (+) nature → churn Partially
H5 Previous churn behavior (+) churn Opposite
H6 Previous churn behavior (+) nature → churn No Table 12: Hypothesis results
Model estimates
As we now know which hypotheses can be confirmed, we can interpret the estimates for the relevant paramters. In the model, both odds ratios and beta’s were estimated. While
32
The odds ratio of recency is 2.058. This means that an increase of recency by 1 standard deviation leads to a decrease in odds of churn by a factor of 2.058. The standard deviation of recency is 80.544. This means that with an additional 80.544 days’ time since the last contact, the odds of churning increase by 105.8%.
Looking at the odds ratio of previous churn behavior, we see that an increase of previous churn behavior by 1 standard deviation leads to a decrease in odds of churn by a factor of 0.287. The standard deviation of years in use is 0.606. This means that with an additional 0.606 churn in the previous years, the odds of churning decrease by 71.3%.
The odds ratio of the interaction effect between frequency and nature is 0.422, meaning that an increase of this effect by 1 would lead to a decrease in the odds of churning by 57.8%. However, both the value of nature and frequency influence the value of the interaction effect. Therefore it is uncertain what the relative impact of each of these individual variables is on the churn probability.
The years in use variable has an odds ratio of 0.218. An increase of years in use by 1 standard deviation leads to a decrease in odds of churn by a factor of 0.218. The standard deviation of years in use is 1.89. This means that with an additional 1.89 years of usage of a method, the odds of churning decrease by 78.2%.
33
Discussion
Conclusion
The goal of this paper was to examine whether customer service contacts and previous churn behavior had any effect on the probability of churning. Several aspects where examined and we can conclude that customer service contacts can, at least partially, predict churn
probabilities. The components of customer service contact that were examined are nature, frequency and recency of customer service contacts. Previous churn behavior was also researched along with moderating effects of previous churn behavior and frequency on the relationship of nature and churn. These components were summarized in 6 hypotheses. Out of all components, recency of customer service contact, previous churn behavior were confirmed to predict churn. An interaction effect of nature and frequency was also confirmed to
somewhat predict churn. In addition, one of the control variables, years in use, was also confirmed to be predictor of churn. To understand what the results mean in the real world, the hypotheses are explained.
H2 was confirmed and thus, we can conclude that there is a negative relationship between recency and churn probability. The results are in line with the expectations stated beforehand. Whatever effects tickets may have, the recency of the tickets matter, most likely due to
involvement of the customer and the ties who are stronger when they have not yet faded away (knox & van Oest, 2014). This indicates that customers who haven’t contacted the company in a while, are more likely to churn.
Frequency doesn’t have a significant effect on churn. This means that the amount of tickets doesn’t necessarily affect churn. Unlike many would expect, the frequency of contacts alone doesn’t seem to matter when it comes to churning.
Nature doesn’t have a significant effect on churn. Some might think that negative tickets would make customers want to churn. This is also supported by literature. However, this is not the case. Negative tickets turn out not to be as bad as many might think. Possible explanations could be that negative tickets either just don’t matter that much, or complaints might be handled well enough that it doesn’t cause customers to churn.
There is a marginally significant interaction effect between nature and frequency. However, both frequency and nature don’t have a direct relationship with churn. This negative
34
of tickets very positive or negative, there is an effect on churn. We must be careful with interpreting this effect since we can’t determine what the relative importance of each of these variable is. One could try to explain the significant effect as follows. When a customer has a large amount of negative tickets, their churn probability could increase. The same reasoning could imply that having a large amount of positive contacts decreases churn probability. We could therefore argue that neither nature or frequency of contacts matters when predicting churn, unless the contact frequency is very high or the nature is overwhelmingly negative or positive. However, we can’t be certain which of these variables is most important in this interaction.
The results of the study showed an opposite effect of H5. Previous churn behavior is negatively related to churn. This means that customers who often churn have a decreased probability of churning. These results contradict the expectation that customers who churn more often, tend to churn repeatedly (Kumar, Leszkiewicz & Herbst, 2018). Switching costs are expected to be lower for the frequent churner, which leads one to expect their churn probability would increase. Although unexpected, this negative relationship can be explained. The sections, who have churned a lot already, know what to expect of a learning product. They also know when a product can be considered good, because they can compare them with past experiences. If the company in this research delivers good products, often churning customers might very well stay because of all products they have tried the current one seems the best.
We confirmed that there is an effect of previous churn behavior on churn probability. We didn’t confirm that nature directly influences churn. Since the interaction effect between nature and previous churn behavior is not significant, we also can’t confirm a moderation effect of previous churn behavior on nature. This tells us that previous churn behavior doesn’t influence the effect of nature of contacts on churn.
35 Managerial implications
Knowing which customers are the most likely to churn enables managers to allocate resources more effectively. Using resources on certain types of customers might be a waste of time and money if their churn probability is low anyway. Allocating resources to customers with higher churn probabilities can allow managers to retain these customers, increasing the effectiveness of these resources.
People who haven’t had recent contact with the company are more likely to churn. Allocating more resources to these customers to try to tie them back to the company could be effective. New customers could also very well be a good segment to allocate more resources to.
Customers who have been using the companies products for a longer time tend not to change their behavior and not churn anyway. New customers need to be convinced more. If a
company manages to retain this customer for a longer time the churn probability will
gradually decrease and the resources allocated to this customer can also be reduced over time. Less resources can be allocated towards customers who have often churned before. These customers tend to settle for products they are satisfied with. Since previous churn behavior is negatively related to churn, it is likely that these frequent churners will stay a customer from the company. Therefore, it is not needed to allocate more resources to this group.
Future research
This research was carried out for the Dutch secondary education market. Primary, higher and professional education where excluded from the research. For future research, it would be interesting to research the effects of customers service contact on churn in these other segments. A major difference between primary education and secondary education is the frequency of switching. Primary schools have a lower budget for these school supplies and therefore usually switch methods every 7 to 8 years instead of the usual 4 years in secondary education. If a school has to use products longer there may be other factors influencing churn or some variables used in this study might have a different importance. One could expect years in use, for example, to have a larger impact due to inertia originating from using products for a very long time. It would be interesting to research these effects.
36
students. Professional and higher education students can decide to either not buy to books or buy secondhand. Both alternatives cause missed revenue for the company. It can be
37
References
Ahn, J. H., Han, S. P., & Lee, Y. S. (2006). Customer churn analysis: Churn determinants and mediation effects of partial defection in the Korean mobile telecommunications service industry. Telecommunications policy, 30(10-11), 552-568.
Alsem, K. J., & Klein Koerkamp, R. (2016). Zorg met liefde en lef: stappenplan voor effectieve zorgmarketing. Bedum: Profiel Uitgeverij.
Athanassopoulos, Antreas D. (2000), “Customer Satisfaction Cues to Support Market Segmentation and Explain Switching Behavior,” Journal of Business Research, 47, 3, 191– 207.
Bell, Simon J., Seigyoung Auh, and Karen Smalley (2005), "Customer Relationship Dynamics: Service Quality and Customer Loyalty in the Context of Varying Levels of Customer Expertise and Switching Costs," Journal of the Academy of Marketing Science, 33(2), 169-83.
Bliese, P. D. (2014, June). Analysis of Multi-Level Organizational Data.
Boles, J. S., Barksdale, H. C., & Johnson, J. T. (1997). Business relationships: an examination of the effects of buyer‐salesperson relationships on customer retention and willingness to refer and recommend. Journal of Business & Industrial Marketing, 12(3/4), 253–264.
Cai, R., & Chi, C. G.-Q. (2018). The impacts of complaint efforts on customer satisfaction and loyalty. The Service Industries Journal, 38(15-16), 1095–1115.
Cheng, L. C., Wu, C.-C., & Chen, C.-Y. (2019). Behavior Analysis of Customer Churn for a Customer Relationship System. Journal of Global Information Management, 27(1), 111–127. Cullen, J. & Shcherbakov, O. (2010), "Measuring Consumer Switching Costs in the Wireless Industry," working paper, Federal Communications Commission.
Doorn, J. V., & Verhoef, P. C. (2008). Critical Incidents and the Impact of Satisfaction on Customer Share. Journal of Marketing, 72(4), 123–142.
Fleiss, J.L. Statistical Methods for Rates and Proportions. New York: John Wiley; 1981. Gelman, A. (2006). Multilevel (Hierarchical) Modeling: What It Can and Cannot
38
Gupta, S., Lehmann, D. R., & Stuart, J. A. (2004). Valuing customers. Journal of Marketing Research, 41(1), 7–18.
Huang, B. Q., Kechadi, T. M., Buckley, B., Kiernan, G., Keogh, E., & Rashid, T. (2010). A new feature set with new window techniques for customer churn prediction in land-line telecommunications. Expert Systems with Applications, 37(5), 3657-3665.
Jahromi, A. T., Stakhovych, S., & Ewing, M. (2014). Managing B2B customer churn, retention and profitability. Industrial Marketing Management, 43(7), 1258–1268.
Koller, M. (2016). robustlmm: An R Package for Robust Estimation of Linear Mixed-Effects Models. Journal of Statistic Software, 75(6).
Knox, G., & van Oest, R. (2014). Customer Complaints and Recovery Effectiveness: A Customer Base Approach. Journal of Marketing, 78(5), 42–57.
Kumar, V., and Werner Reinartz. Customer Relationship Management: Concept, Strategy, and Tools. Berlin: Springer Berlin, 2012.
Kumar, V., Leszkiewicz, A., & Herbst, A. (2018). Are you Back for Good or Still Shopping Around? Investigating Customers Repeat Churn Behavior. Journal of Marketing
Research, 55(2), 208–225.
Kumar, V., & Shah, D. (2009). Expanding the Role of Marketing: From Customer Equity to Market Capitalization. Journal of Marketing, 73(6), 119–136.
Landis, J.R., Koch, G.G. “The measurement of observer agreement for categorical data.” Biometrics. 1977;33 (1):159-174.
Leeflang, Peter S. H., Jaap E. Wieringa, Tammo H. A. Bijmolt, and Koen H.
Pauwels. Advanced Methods for Modeling Markets. Cham: Springer International Publishing, 2017.
Leeflang, Peter S. H., Jaap E. Wieringa, Tammo H.A. Bijmolt, and Koen H.
Pauwels. Modeling Markets: Analyzing Marketing Phenomena and Improving Marketing Decision Making. New York: Springer, 2015.
Lilien, G. L. (2016). The B2B Knowledge Gap. International Journal of Research in Marketing, 33(3), 543–556.
39
Rauyruen, P., & Miller, K. E. (2007). Relationship quality as a predictor of B2B customer loyalty. Journal of Business Research, 60(1), 21–31.
Retana, G. F., Forman, C., & Wu, D. J. (2016). Proactive Customer Education, Customer Retention, and Demand for Technology Support: Evidence from a Field
Experiment. Manufacturing & Service Operations Management, 18(1), 34–50.
Risselada, Hans, Peter C. Verhoef, and Tammo H.a. Bijmolt. “Staying Power of Churn Prediction Models.” Journal of Interactive Marketing 24, no. 3 (2010): 198–208. https://doi.org/10.1016/j.intmar.2010.04.002.
Rust, R. T., & Cooil, B. (1994). Reliability Measures for Qualitative Data: Theory and Implications. Journal of Marketing Research, 31(1), 1–14.
Slack, N. J., & Singh, G. (2020). The effect of service quality on customer satisfaction and loyalty and the mediating role of customer satisfaction. The TQM Journal, 32(3), 543–558. Umashankar, N., Ward, M. K., & Dahl, D. W. (2017). The Benefit of Becoming Friends: Complaining after Service Failures Leads Customers with Strong Ties to Increase
Loyalty. Journal of Marketing, 81(6), 79–98.
Vincent, O. R., Makinde, A. S., & Akinwale, A. T. (2017). A cognitive buying decision-making process in B2B e-commerce using Analytic-MLP. Electronic Commerce Research and Applications, 25, 59–69.
