The effect of promotional
behaviour on customer churn
The effect of promotional behaviour on customer churn
Raymond Kroes
University of Groningen Faculty of Economic and Business
Master thesis for MSc Marketing June, 2014
Nieuwe Boteringestraat 11A1 9712PE Groningen
06 83701639
r.kroes@student.rug.nl
Student number 1697986
Management summary
In order to obtain positive financial outcomes, companies should devote to decrease customer churn or increase sales, for instance by offering promotions. Prior research found evidence for effects of several determinants on customers churn, but little is known about the effect of promotional behaviour. This study aims to assess the relationship between promotional behaviour and customer churn.
A literature search was conducted to identify determinants that contribute to the likelihood of a customer to churn, including switching costs, customer dissatisfaction, length of a relationship, frequency, recency, monetary value and customer characteristics. In addition to these determinants, this study proposes three hypotheses about promotional behaviour that constitutes to a customer’s likelihood to churn. A dataset provided by of an online greeting card company is used in order to develop a model and test the hypotheses. The dataset contains information about more than 150,000 Dutch customers and has been obtained between August 2011 and July 2012.
Results indicate that whether a customer receives a discount on his first purchase as well as the number of discounts a customer receives, has a significant negative impact on the probability to churn. The value of the received discount as a percentage of gross sales on the other hand, has a significant positive impact on the probability to churn. Therefore, in order to decrease a customer’s probability to churn, companies should offer customers a discount on their first purchase, offer discounts more often and should not offer too high discounts. Furthermore, the percentage of discount is most important of all three variables in predicting churn behaviour.
All three variables are found to have a significant effect on the probability to purchase in the future again. Interestingly however, whether a customer receives a discount on his first purchase does not have an opposing sign in contrast to the probability to churn. The variable has a negative sign in both situations, which might be caused by the time that passes. The number of discounts a customer receives and the value of the received discount as a percentage of gross sales do have opposing signs.
Table of contents
List of figures ... 3
List of tables ... 3
1. Introduction ... 4
1.1 Background and research problem ... 4
1.2 Purpose of the study ... 5
1.3 Methodology ... 6
1.4 Outline ... 7
2. Literature review ... 8
2.1 Customer churn behaviour ... 8
2.2 Determinants of Churn ... 9
2.3 An online environment ... 20
2.4 Conceptual model ... 20
3. Research data ... 22
3.1 Research data ... 22
3.2 Missing values and outliers ... 23
4. Methodology ... 25
4.1 Classification trees ... 25
4.2 Binary logit model ... 25
4.3 Cox regression model ... 27
5. Results ... 29
5.1 Classification trees ... 29
5.2 Binary logit model ... 31
5.3 Cox regression model ... 40
6. Discussion and conclusion ... 43
6.1 Implications ... 43
6.2 Limitations and further research ... 45
References ... 47
3
List of figures
Figure 1: The effect of promotional behaviour and other determinants on the probability to
churn... 21
Figure 2: Cumulative lift curve with Tree models ... 30
Figure 3: Cumulative lift curve with a Binary logit model ... 34
Figure 4: Effect of First purchase discount on predicted probability to churn ... 36
Figure 5: Effect of Percentage of discount on predicted probability to churn ... 37
Figure 6: Effect of Number of discounts on predicted probability to churn... 38
Figure 7: Survival function of predicted probability to purchase ... 41
List of tables
Table 1: Independent variables and their descriptions ... 22Table 2: Different months of churning ... 32
Table 3: Results of the Binary logit model ... 33
Table 4: Outcomes of hypotheses ... 38
Table 5: Results of the Cox regression model ... 41
Table A1: Means first purchase discount………..54
Table A2: Means percentage of discount………..………..54
Table A3: Means number of discounts………..54
4
1. Introduction
1.1 Background and research problem
Companies rely considerably on the retention of customers to obtain positive financial outcomes in the long-run (Anderson and Mittal, 2000). A lack in retention of customers is costly for two reasons: firstly, a decline in number of customers is associated with revenue loss. Secondly, the cost of acquiring new customers is higher than that of keeping the same (Reichheld and Sasser, 1990; Reinartz, Thomas and Kumar, 2005). Therefore, companies should devote to increasing customer loyalty and decreasing customer churn.
Another method to obtain positive financial outcomes is to increase the company’s sales of products or services. There are several options to achieve this. One option is the promotion of sales by offering a lower price for a limited period of time. Numerous researchers have identified the positive effect of purchase promotions on purchase intentions (e.g. Gabler and Reynolds, 2013) or sales (e.g., Moriarty, 1985; Ataman, Van Heerde and Mela, 2010; Ikenna Ofoegbu and Mfonobong Udom, 2013).
Although several determinants of customer churn behaviour have been identified in prior studies (e.g. Schmittlein and Peterson,1994; Moe and Fader, 2001; Ahm, Han and Lee, 2006), and the effect of promotions on sales has been investigated intensively, the effect of promotions on customer churn behaviour remains unclear. However, there are indications for a possible relationship between these variables.
5
Therefore, this study aims to answer the following research question:
What is the effect of a customer’s promotional behaviour on a customer’s probability to churn?
1.2 Purpose of the study
Although most research about promotional behaviour focuses on the effect of a consumer’s purchase intentions and sales, little is known about the effect of price promotions on churn behaviour. The purpose of this study is to contribute to the existing literature of churn behaviour by assessing the role of promotional behaviour. Through conducting an empirical study with the main focus on promotional behaviour, new insights could be obtained about the importance of price promotions on a customer’s likelihood to churn. The effect of promotional behaviour on churn can be affected by the following factors: whether a consumer purchases his or her first purchase on a discount; how much percentage of discount a consumer receives; and how many times a consumer purchases a product on a discount. To increase the predictive validity of churn models, it is important for researchers to know if there is a relationship between promotional behaviour and customer churn behaviour; what the direction is; and how strong this relationship is. By empirically developing a model that can predict churn, insights may be obtained regarding the importance of different determinants when variables for promotional behaviour are included.
6
customers are more likely to churn, companies could react appropriately with customised marketing strategies and incentives to encourage customers to stay. In addition, it may be interesting for managers to know what the optimal amount of discount is: will more or less discount be better or is there an optimal amount of discount in between? This is particularly important, as a higher amount of discount might cost the company a significant amount of money. Neslin, Gupta, Kamakura, Lu and Mason (2006) report that churn prediction models last for at least 3 months. Consequently, it would not be required to develop a model every month, saving companies a considerable amount of money and effort. IMoreover, models that could predict customer churn behaviour accurately are meaningful from a managerial perspective, since it could significantly increase profit.
1.3 Methodology
Data used
In this study, a dataset provided of an online greeting card company is used. The dataset contains information about more than 150.000 Dutch customers and has been obtained between August 2011 and July 2012. The dataset provides information about churning behaviour, discounts, the number of complaints, relationship duration, number of orders, other customer related characteristics, and behaviour for each individual customer. The dataset will be described in more detail in chapter 3.
Research methods used
The study can be divided into two parts. Firstly, a literature review will give insights in different determinants which were found in prior research. Secondly, in a statistical analysis three models are developed to predict the probability to churn and test the hypotheses.
7
independent variables and the probability to churn as a dependent variable. The methodology of this study will be explained in more detail in chapter 4.
1.4 Outline
8
2. Literature review
The purpose of this chapter is to discuss the concept of customer churn behaviour and promotional behaviour. In addition, the concepts of different determinants that affect customer churn behaviour will be discussed. The relationship between these determinants and the likelihood of a customer to churn will also be described. This chapter will conclude with the introduction of the conceptual model of this study.
2.1 Customer churn behaviour
Neslin et. al (2006) define churn as the tendency of customers to end a relationship with a company in a given period of time. Customer churn is related to a direct reduction of income and an increasein numerous additional costs (Reichheld and Sasser, 1990). Most important are the costs associated with acquiring new customers in order to maintain the level of customer base. According to Reinartz, Thomas and Kumar (2005), if a company has to decide whether it will have a suboptimal allocation of acquisition expenditures or retention expenditures, a suboptimal allocation of retention expenditures will have a greater impact on long-term customer profitability than a suboptimal allocation of acquisition expenditures. Hence, reducing customer churn by targeting customers with the highest probability to churn with customised designed campaigns will be more effective than acquiring new customers (Verbraken, Verbeke and Baesens, 2014).
9
In spite of its importance, churning remains a serious issue across different industries, including the internet, insurance and cable service industries. Churn rates appear to be rather similar in countries and over time (OECD, 2013). On average, 15-25% of all customers in the services industry and 10-20% of all customers in the manufacturing industry exhibit churn behaviour within a year (OECD, 2013). When separating the service market in different submarkets, the telecommunication service market appears to have the highest churn rate, with an average of 20-40% per year (Keramati and Ardabili 2011).
In order to measure the customer churn rate, companies have to divide the ‘number of existing customers who have left by the end of a given period by the number of existing customers at the beginning of the respective period’ (Kumar and Reinartz, 2012). The resulting figure is a percentage that represents the relative number of customers who end the relationship with a company. However, it does not account for the likelihood of individual customers to churn. In order to be able to manage customer churn, companies should aim to understand a customer’s churn behaviour and the determinants that cause churn.
2.2 Determinants of Churn
Various determinants of customer churning behaviour have been found in prior studies. Although these determinants were found in different industries and for different products than the data of the current study, the OECD (2013) found that the underlying causes of churn are similar across industries and products and that the level of churn rates appear to be similar in countries and over time as well.
10
Bawa and Shoemaker (1987), Gabler and Reynolds (2013), and others): the number of discounts per customer, whether the first purchase involved a discount, and the value of the total discount received as a percentage of gross expenses.
In order to develop a complete model with high predicative validity, other variables will be included as well. Schmittlein and Peterson (1994) report that past purchase behaviour is a significant positive determinant of future purchasing behaviour. This is the result of inertia, which implies that purchase behaviour reacts to external stimuli with a certain slowness (Brown, 1952). A lack of past purchasing therefore implies a higher probability of low purchase behaviour in the future. According to Anderson, Weitz (1989) and McDonald (2010), the length of a relationship is an important determinant of future stability of a relationship. In a study about customer churn in the South Korean mobile telecommunications service market, Ahm, Han and Lee (2006) identified evidence for different determinants of customer churn. The amount of money spent on a product or service (i.e. monetary value), customer dissatisfaction and switching costs proved to be most important. These causes of customer churn are confirmed in studies by Matthews and Murray (2006) and Svendsen and Prebensen (2013). According to Kumar and Reinartz (2012) and others (e.g. Moe and Fader, 2001), recency, frequency and monetary (RFM) values are used to classify customers in order to evaluate customer behaviour and customer value. According to Wu and Zheng (2005) it has been proven that recency, frequency and monetary value are very effective in predicting customer behaviour when applied to marketing databases. Buckinx and Van den Poel (2005) report that recency, frequency and monetary values are better used in predicting behavioural loyalty than demographics. Seo, Ranganathan and Babad (2008) concompany that demographical variables have a weak, but significant relationship with churn. Age, gender and region will therefore be included as well.
2.2.1 Promotional behaviour
11
different products, prices of the same product at different retailers can be compared relatively easily as well. Price is a key element in any market exchange and can represent the amount of money a customer is willing to pay for a certain product or service (Lichtenstein, Ridgway, & Netemeyer, 1993). Although some consumers might perceive that a too low price implies lacking product quality, other consumers will buy products for a price as low as possible (price-sensitive consumers). For these customers, lower prices are the main explanation of their purchases (Lichtenstein, Ridgway, & Netemeyer, 1993).
A reference price is closely related to the price. A reference price is the subjective price level with which consumers compare observed prices (Alvarez and Casielles, 2005). Reference prices are prices relative to what consumers paid in the past and what they expect to pay in the future (Gabler and Reynolds, 2013), and are related to the comparison between the observed price (i.e. the discount price) and the original price (Kalwani, Yim, Rinne and Sugita, 1990). In the latter case, consumers exhibit promotional behaviour. The importance of the different aspects of promotional behaviour on churn is described below.
2.2.1.1 First purchase discount
12
not price-sensitive. Price-sensitivity will therefore have a direct effect on consumer’s switching behaviour during the acquisition phase. Consequently, the following hypothesis is proposed:
H1 Whether a customer receives a discount with his or her first purchase, has a positive effect
on a customer’s probability to churn.
2.2.1.2 Percentage of discount
In the definition of price-sensitivity suggested by Abad and Jaggi (2003), the sensitivity of the demand of a customer depends on the observed selling price (i.e. the price charged by the retailer). However, this definition does not contain a relative conception and therefore does not incorporate a comparison with prices of similar products or the focal product with a discount price. Consumers might react differently to a comparison between a discount price and its normal price than they would react to a comparison between the price of product A and the price of product B. Moreover, ‘consumers view prices relative to what they paid in the past and what they expect to pay in the future’ and therefore ‘create their own reference price or standard by which they judge the price of an item’ (Gabler and Reynolds, 2013, p. 442). This suggests that not only during the acquisition phase, but also during the past and expected future relationship, prices have an important role. The question rises whether customers react similarly to prices with discounts as to prices that are relatively low compared to products from other brands (see chapter 2.2.1.1). Do customers who receive higher percentages of discounts have a higher likelihood to exhibit switching behaviour than customers who purchase with less discounts? The following hypothesis is proposed:
H2 The extent to which the value of the total discount received as a percentage of gross sales
is higher, has a positive effect on a customer’s probability to churn.
2.2.1.3 Number of discounts
13
This is in line with findings of Gabler and Reynolds (2013), who report a positive effect of the number of discounts on purchase intentions for highly visible products.
Although the number of discounts have a positive relationship with purchase intentions and purchase behaviours, the effect of the number of discounts on churn remains unclear. However, a churned customer implies that he will not purchase any products in the future anymore. Therefore, churning could be considered as the opposite of purchasing. Consequently, higher number of discounts results in a higher likelihood to purchase and suggests a lower likelihood to churn. Therefore, the following hypothesis is proposed:
H3 The extent to which a customer receives discounts more often, has a negative effect on a
customer’s probability to churn.
2.2.2 Switching costs
Customers continue relationships with companies for two reasons: they either stay because of loyalty or because they have to make an effort in order to change from one company to another (Bendapudi & Berry, 1997). These efforts can be expressed in terms of costs: switching costs. Switching costs describe all variety of costs (both financial and non-financial) experienced in changing suppliers and is the perception of the amount of additional expenses associated with leaving a company compared to those of staying with a company (Matthews and Murray, 2006; Patterson & Smith, 2003). By building specific tangible and intangible assets, companies can build long term relationships with customers. This will increase customers’ switching costs. Moreover, as the duration of a relationship between a customer and a company increases, the amount of customers’ trust in a company’s offerings also increases (Coulter and Coulter, 2002). The quality of these offerings is related to customers’ perception of the reliability and competences of a company. The increase in this perception will, in turn, result in an increase of switching costs (Aydin and Ozer, 2005).
14
explains why the switching costs for a customer should not be too high, in order to switch to another company. Switching costs therefore have a negative effect on customers’ probability to churn.
Loyalty points are, next to membership card programs, a major component of switching costs, since all the accumulated loyalty points may be lost when customers switch from one company towards another (Ahm, Han and Lee, 2006). In order to maintain their loyalty points, customers have a lower tendency to switch.
2.2.3 Customer dissatisfaction
Customer satisfaction is a commonly applied metric in marketing and is defined as the perception of pleasant fulfillment of a service (Oliver, 1997). More specifically, customer satisfaction is the resulting feeling of consumers when making a positive evaluation or feeling happy about their decision (Hoyer, MacInnis and Pieters, 2013). However, if customers sense a negative evaluation of an outcome, they will become dissatisfied. Dissatisfaction is the result of a discrepancy between expectations and performance (Hoyer, MacInnis and Pieters, 2013). So in order to satisfy customers, companies should not create too large expectations and make sure that their performance is higher than the expectations of customers. However, if companies do not succeed, customers will become dissatisfied.
Companies will only know whether customers are dissatisfied if customers communicate their criticism to the company, communicate their criticism in public or stop doing business with the company. On the other hand, dissatisfied customers may also do nothing or communicate their criticisms to their friends, family members and colleagues. Although Day (1977) reports that most dissatisfied consumers do not complain, it is difficult to determine the actual number of complaining customers, because in addition to the company, consumers complain, also to other consumers or in public. In an online environment, complains can be spread relatively fast and easily, compared with an offline environment. Therefore, complaining customers indicate problems that require attention.
15
identify a weak relationship (Ahm, Han and Lee, 2006; Seo, Ranganathan and Babad, 2008). On the other hand, there are also studies that attempt to find the effect of the intention to churn or perceived (non)loyalty on attitudinal loyalty. These studies find high explanatory values, up to R2=0,85 (Xiaoyun, Kwortnik and Wang, 2008). However, the actual behaviour of customers has not been incorporated (i.e. the actual probability to churn) in the studied model.
Solnick and Hemenway (1992) researched complaining customers and report that complaints are usually based on consumer recollections instead of actual complaints. They noticed a positive relationship between the number of complaints registered and the customer churn rate. A complaint of a customer and exiting often occur together, so can be considered complementary.
2.2.4 Length of relationship
One way to categorise customers in a company’s customer base could be by the length of their relationship with the company. To define the length of a relationship, a distinction should be made between attitudinal loyalty and behavioural loyalty. According to the CUSAMS-framework of Bolton, Lemon and Verhoef (2004), attitudinal loyalty is a consumer’s satisfaction, commitment and price perception and will affect behavioural loyalty. Along with depth and breadth of a relationship, the length of the relationship is part of consumers’ behavioural loyalty. Consumers that have a long relationship with a retailer are more behaviourally loyal than consumers that have a short relationship or no relationship at all. Hence, behavioural loyalty is defined as the repurchase of a brand of interest (Ehrenberg, 2000).
16
the relationship. Therefore, both behavioural as well as attitudinal loyalty are influenced by the length of the relationship between a customer and a company.
During the relationship, customers may develop an understanding of a company’s procedures and may become familiar with interacting with the company. As the length of a relationship between a customer and a company increases, customers may develop a higher perceived identification with the company (Bhattacharya, 1998). As demonstrated in the CUSAMS-framework (Bolton, Lemon and Verhoef, 2004), customers who could identify themselves better with a company (i.e. those that are committed to a company), will show increased attitudinal loyalty, which will have a direct effect on the length of a relationship and therefore on behavioural loyalty as well. In turn, those customers that are behaviourally loyal to a company are less tended to switch to another company.
In a 2010 study on churn rates amongst season ticket holders of professional sport teams, McDonald reports that the length of a relationship is, in addition to the number of games attended, the key factor in predicting churn. This is supported by Andersons and Weitzs’ (1989) findings, who established that the future stability of a relationship between a customer and a company is positively influenced by the length of a relationship.
2.2.5 Recency
According to Blattberg, Byung-Do and Neslin (2008) recency belongs, in addition to frequency and monetary value, to one of the pillars of customer relationship management. Recency is defined by the time that has passed since a customer’s previous purchase (Neslin, Taylor, Grantham and McNeil, 2013). Therefore, a higher recency means that a longer period of time has passed since a customer’s last purchase.
Various studies have explained the effect of recency on predicting the probability of purchase behaviour and established that there is a high correlation between both variables; a higher recency is associated with a lower probability to purchase (e.g. Fader, Hardie and Lee, 2005). This is in line with the findings of Wu and Chen (2000), who report that customers who purchased recently have a higher probability of being active than customers who did not purchase recently.
17
According to the authors, companies face a “recency trap” when customers do not purchase for a long period of time. The recency will increase, which makes it less likely that customers will purchase in the next period and therefore increases the recency of purchasing of that period, etcetera. As a result, customers are slightly “floating away” from the company and in the end will not purchase any products.
The value of recency does not only have a direct effect on the behaviour of customers, but also on the perception of customers. If the value of recency is low, the likelihood of a consumer being attitudinally loyal is increased (Buckinx and Van den Poel, 2005). As mentioned before, the CUSAMS-framework (Bolton, Lemon and Verhoef, 2004) explains that attitudinal loyalty has an effect on behavioural loyalty, which in turn implies that customers are less tended to churn. In a non-contractual setting, there is evidence that recency is the most important determinant in order to predict whether the relationship between a customer and a company is active or inactive (Reinartz and Kumar, 2000).
2.2.6 Frequency
Several studies established the positive relationship between the frequency of purchases and a change in customers’ future behaviour (e.g. Schmittlein and Peterson, 1994; Lemon, White and Winer, 2002). Frequency is a commonly used measurement to predict customer behaviour, because frequency is positively related to customers’ expected use of products in the future (Lemon, White and Winer, 2002). Reinartz and Kumar (2012) define frequency as ‘a measure of how often customers order from the company in a certain defined period’. Therefore, if a customer did not acquire many products in the past, it is expected that he or she will not purchase many products in the future and in the end will float away from the company. Those consumers therefore have a higher probability to become a churning customer.
18
both offline as well as online shopping situations is, among other factors, influenced by the frequency of purchases.
Several studies report that, on average, consumers who visit an e-commerce website or traditional store more often have a higher probability of purchasing (Moe and Fader, 2001). In addition to frequency, the evolution of visiting an e-commerce website is related to the probability of purchasing as well. Consumers who visit a store at an increasingly frequent rate also have a higher probability of purchasing than consumers with a decreasing visiting frequency (Moe and Fader, 2001). A combination of visit frequency and an increase in the evolution of visit frequency leads to the highest probability of purchasing. Therefore, to increase selling and reduce churn, retailers should increase both the visit and purchasing frequency of consumers.
2.2.7 Monetary value
In addition to recency and frequency, the monetary value is part of the RFM model. This is a frequently used measurement method which is introduced in the introduction of this chapter. The underlying concept of the monetary value in the RFM model, is that customers who spent more money, have a higher probability to buy again than those who spent less money. Therefore, the total amount of money someone spends at a company will be represented by the monetary value (Buckinx and Van den Poel, 2005). Reinartz and Kumar (2012) also take the number of products into account and define the monetary value as the average amount that a customer spends on a transaction. In the present study, the definition of Buckinx and Van den Poel will be used, because the number of purchases is taken into account in the frequency of purchases.
19
belief that consumers are rational and will choose the product with the lowest price if all other product attributes are equal. In addition, Ahm, Han and Lee (2006) support the hypothesis that heavy users are more likely to churn. They also suggest that heavy users have the tendency to intensively explore more other brands with better product attributes.
2.2.8 Age, gender and region
In addition to customer behavioural variables, other variables have been reported in the literature that might affect the probability of churn are of a demographic nature, including: age, gender and region. Seo, Ranganathan and Babad (2008) report that age and gender are significant, but weak determinants of churn in the telecommunication service market. The reason for these significant differences between males and females and different ages, is that there are differences in the usage of their mobile phone. This is in line with the findings of Buckinx and Van den Poel (2005) in a study about the fast moving consumer good environment. They report that recency, frequency and monetary values are better factors in predicting behavioural loyalty than demographics. Svendsen and Prebensen (2013) however, identified no significant relationship of gender on churn and a significant, but weak negative relationship of age on churn. Older people might have longer relationships with companies or are less able to compare alternatives. The findings of Svendsen and Prebensen (2013) concompany this statement, which implies that older people have a smaller likelihood to churn.
20
2.3 An online environment
Due to the introduction and increased use of the internet, the economic marketplace of supply and demand has changed entirely. Instead of being limited to visiting the nearest store in the neighbourhood at limited opening hours, customers have become able to shop wherever and whenever they want. This evolution causes a change in the customer base of a company. Instead of having relations with customers who live close to the physical location of a company, companies are now able to serve customers in a large geographical area. However, it is easier for customers to switch to other companies or spread their spending among companies, since the online environment offers a variety of alternatives.
According to the OECD (2013) the level of churn rate appears to be similar in countries and over time. Therefore, it is assumed that the underlying causes of churn will be similar across industries and products as well. Because of the recency of the OECD report (2013), both offline as well as online environments were taken into account. So, despite the differences between an offline and an online environment mentioned above, the determinants of churn are assumed to be similar.
2.4 Conceptual model
21
Figure 1: The effect of promotional behaviour and other determinants on the probability to churn. H1, H2 and H3 are the hypotheses proposed in the literature review.
22
3. Research data
The purpose of this chapter is to introduce the research data used for this study. Firstly, a description of the data used in this study is described and an overview of all variables from the conceptual model is provided. Secondly, the missing values and outliers in the dataset are described and removed.
3.1 Research data
In this study, a dataset provided of an online greeting card company is used. The dataset contains information about 138,545 Dutch customers and has been obtained between August 2011 and July 2012. The data contains information, among other, about churning behaviour, promotional behaviour, complaints, recency, frequency and monetary value. A major advantage of revealed data compared to stated data, is that it is derived from real transactions and that it measures real consumer behaviour. Therefore, it has higher validity than stated data (Huang, Haab and Whitehead, 1997). Table 1 presents the names, levels and descriptions of the determinants (independent variables).
Table 1: Independent variables and their descriptions
Determinants Description Levels
First purchase discount
First order voucher code 1=yes, 0=no Percentage of
discount
Value of the total discount received as a percentage of gross sales
0.00-1.00 (0%-100%) Number of discounts Number of orders with voucher ≥0
Switching costs Total number of loyalty points in spite of being used or not being used
≥0 Customer
dissatisfaction
Number of complaints registered ≥0 Relationship length Number of months since the date that the customer
registered as a customer
≥0 Recency Number of months since the date that the customer placed
his last order
0-11 (with 11=one year ago, 0=last month)
Frequency Total number of orders placed by the customer ≥1 Monetary value Total sales value from the customer, including discount value
in €
≥0.00
Age Age in years at 07-2012 10-100
Gender The customer´s gender 1=male, 0=female Location Number of citizens per postal code ≥1
23
of customers. Most customers from the dataset purchase a product only once, twice or three times a year (approximately 11.2% of all customers). A period of 13 months includes the same date twice, so includes those customers who purchase only once a year and determines whether they purchase a year later again. This assumption is further investigated in chapter 5.2.
3.2 Missing values and outliers
There are 15,200 customers in the dataset that have one or more missing values. The summation below presents the variables containing missing values and possible explanations.
Location: Customers that live outside the Netherlands or have an unknown address
are not taken into account. This variable accounts for most of the missing values.
Age: Customers who filled in that they were younger than 10 years of age or older
than 100 years are considered to be unreliable and are not taken into account.
First purchase discount: For some customers it is unknown whether they bought
their first product on a discount or not.
Eight outliers have been detected (with the use of boxplots) for the following variables:
Frequency: Customer ID 28696973 (a value of 1,108);
Number of discounts: Customer ID 28548306 (a value of 193), 29413142 (a value of
212) and 257488767 (a value of 224);
Monetary value: Customer ID 27947232 (a value of € 6,275.07), 29172169 (a value of
€ 8,176.26), 613175256 (a value of € 9,276.06) and 29413142 (a value of € 10,075.78).
After removing all customers that have one or more missing values or an outlier, 138,545 customers remain in the dataset.
24
25
4. Methodology
The purpose of this chapter is to describe the analyses used in this study. This chapter is subdivided in the following methods: classification trees, a Binary logit model and a Cox regression analysis.
4.1 Classification trees
In order to examine whether the determinants introduced in chapter 2 have an effect on customer churn behaviour, and which of these determinants are most important, a classification tree is developed. A classification tree is a commonly used method to predict churn, because it predicts well on large datasets, and generates interpretable models (e.g. Nesling et al., 2006; Risselada, Verhoef and Bijmolt, 2010). A decision tree consists of four key elements: a Root node, Child nodes, Terminal nodes and Splitting rules. The ultimate goal is to split nodes (which are based on the determinants) in order to optimally predict churn behaviour. A key benefit of trees over alternative methods such as logistic regression is that it does not make assumptions regarding the distribution of the data, so does not require normally distributed variables.
In order to determine the point(s) where the independent variables have to split, several splitting rules could be used, e.g. CHAID, Exhaustive CHAID and CART. The main difference between these rules is the method used to define a good fit. Haugthon and Oulabi (1993) report that in defining the best splitting rule one has to perform both CHAID and CART and compare them with each other. Therefore, this study will compare the outcomes of different classification trees in order to find consistent results across trees. Those determinants that havethe largest effect on churn according to the decision trees will be used in further analyses.
4.2 Binary logit model
26
regression is conceptually simple and therefore relatively easy to interpret (Bucklin and Gupta, 1992) and contains the availability of a closed-form solution for posterior probabilities (Coussement and Van den Poel, 2008).
A logistic regression model uses a cumulative logistic distribution to transform a linear probability model such that the probabilities follow a s-shaped curve, resulting in interpretable probabilities (i.e. probabilities >1 or <0, as in a linear probability model). The following model is proposed (assuming that, after performing the classification trees, all determinants are included):
Where,
= Predicted probability to churn = Constant
i = Customer ID
FDi = First purchase discount
PDi = Percentage of discount NDi = Number of discounts SCi = Switching costs CDi = Customer dissatisfaction RLi = Relationship length Ri = Recency Fi = Frequency MVi = Monetary value Ai = Age Gi = Gender Li = Location
The primary purpose of the model is to predict churn behaviour and to test the hypotheses. A robustness check is done to determine whether 13 months is a good assumption for customers being considered as churned. In the robustness check, a comparison is made between models with different number of months as dependent variable. The purpose is to examine whether the findings are not sensitive to the definition of the dependent variable and will hold for shorter or longer time periods as well.
27
informative about the models generalisability (Leeflang, Wittink, Wedel and Naert, 2000). Since the model is populated with cross-sectional data, the dataset is randomly split into two samples. The first sample is an estimation sample (which contains 95% random sampled customers) and the second sample a validation sample (which contains 5% random sampled customers).
4.3 Cox regression model
The hypotheses introduced in chapter 2 proposed significant effects of First purchase discount, Percentage of discount and Number of discounts on a customer’s probability to churn. Because the dataset contains non-contractual customers and it is therefore not clear whether they will churn in the future (after April 2014), the dependent variable is set to a concrete time period to determine whether customers are considered as being churned (see chapter 3.1). However, it is known which customers did purchase products between August 2012 and April 2014. Therefore, in addition to testing whether promotional behaviour has an influence on a customer’s likelihood to churn, a Cox regression analysis is used to determine whether promotional behaviour has an effect on a consumer’s future purchase behaviour.
A churned customer implies that he will not purchase any products in the future anymore. Therefore, churning could be considered as the opposite of purchasing. In line with this reasoning, promotional behaviour should show inverse results on the probability to purchase in the future compared to on the probability to churn.
28
Where,
= Estimated probability that a customer purchases at time t, given that he did not purchase until time t
= Baseline hazard at time t
29
5. Results
This chapter will describe the results of the analyses. In order to obtain preliminary insights in the importance of the determinants, a classification tree will be presented first. This study will continue with those variables that areimportant in predicting churn behaviour according to the classification trees,. To test the hypotheses of this study, a Binary logit model is presented. This chapter will finish with a Cox regression model, in order to determine whether promotional behaviour has an influence on a consumer’s future purchase behaviour.
5.1 Classification trees
To obtain preliminary insights in the importance of the different determinants of churning behaviour, several decision trees are developed, with the use of different splitting rules (i.e. CHAID, Exhaustive CHAID and CART). Since the purpose of tree models is obtaining insights, a well-organised model should be developed. Therefore, the parent node is set equal to 100 and the child node to 50 (which is the minimum size; other parent nodes and child nodes are also tested and show similar predicting results). Different splitting rules show comparable results of Hit rates (Hit rate of CHAID = 87.4, Exhaustive CHAID = 87.3 and CART = 87.4) and Top-decile lifts (Top-decile lift of CHAID = 4.17, Exhaustive CHAID = 4.17 and CART = 4.27).
30
Figure 2: Cumulative lift curve with Tree models
All three models use similar determinants to split the nodes. The following determinants are, according to all trees, the most important predictors of customer churn behaviour and end high up in the tree: Frequency, Percentage of discount, Monetary Value, Recency, Relationship length and Switching costs.
The importance of Switching costs seems intuitive, because loyalty points indirectly include frequency and the monetary value as well. Potential cause of multicollinearity issues will be further investigated in chapter 5.2.4.4. As mentioned above, Percentage of discount is an important variable. However, First purchase discount and Number of discounts are less important variables.
31
5.2 Binary logit model
As mentioned in chapter 3, the dataset is divided into an estimation sample and a validation sample. The estimation sample is used to develop the model. The validation sample is used to determine the model’s predictive validity. But firstly, a robustness check on the dependent variable is performed.
5.2.1 Robustness check
In order to check the performance of the variable that indicates whether a customer is considered as being churned (i.e. the dependent variable), a robustness check is performed by changing 13 months into 11, 12, 14 and 15 months. The main purpose is to examine whether the findings are not sensitive to the definition of the dependent variable and will hold for shorter/longer time periods as well. A comparison between different dependent variables is judged based on the significance, signs and size of the parameter estimates.
32
Table 2: Different months of churning (* p<.05; ** p<.10)
Considered as being churned after nr of months: 11 12 13 14 15
Variable Beta Beta Beta Beta Beta
(constant) -.518* -.601* -.658* -.705* -.751*
Promotional behaviour:
First purchase discount -.062* -.071* -.078* -.076* -.071*
Percentage of discount .173* .243* .265* .287* .330* Number of discounts -.048* -.049* -.049* -.050* -.055* Switching costs .008* .009* .009* .010* .010* Customer dissatisfaction -.023** -.027* -.025** -.023** -.027* Relationship length .011* .011* .011* .011* .010* Recency .068* .071* .069* .069* .068* Frequency -.149* -.154* -.156* -.158* -.157* Monetary value -.003* -.003* -.003* -.003* -.003* Age .013* .013* .014* .014* .015* Gender -.070* -.071* -.065* -.063* -.060* Location .000* .000* .000* .000* .000*
5.2.3 Evaluation of the overall model
The Binary logit model is overall significant (Likelihood ratio/χ2
= 41349.350; p-value = 0.00). In order to evaluate how well the model fits, the Cox & Snell R2 and Nagelkerke R2 are used. Both R2’s are “pseudo-R” statistics and should be interpreted differently from the R2 since it is an approximation of the R2 and measure the fit compared to a null model (a model which only includes a constant). The higher the value, the higher the fit of the model (Cox & Snell R2 on a scale from 0-0.75 and Nagelkerke R2 on a scale from 0-1 (Cohen, Cohen, West and Aiken, 2002)). The developed model has a Cox & Snell R2 of 0.270 and a Nagelkerke R2 of 0.428, which mean that the model explains approximately 36-43% (Cox & Snell R2 on a scale from 0-1: 0.270/0.75 = 0.36; Nagelkerke R2 = 0.428) more of the variance in churn than a null model would have.
33
Customer Dissatisfaction has in this study a negative effect on churn, which implies that if a customer complains more often, he or she is less likely to churn. The reason behind this is not clear. This effect might be explained by an effective complaint management policy by a company. Consumers who are customer for a longer period of time are more likely to churn than those consumers who are customer for a shorter period of time. This effect might be explained by customers becoming bored with the assortment, wanting to try something else.
Table 3: Results of the Binary logit model
Variable Beta P-value Exp(Beta) VIF
(constant) -.658 .000 .518
Promotional behaviour:
First purchase discount -.078 .000 .925 1.148
Percentage of discount .265 .000 1.303 1.222 Number of discounts -.049 .000 .952 2.358 Switching costs .009 .000 1.009 3.397 Customer dissatisfaction -.025 .060 .976 1.147 Relationship length .011 .000 1.011 1.123 Recency .069 .000 1.072 1.283 Frequency -.156 .000 .856 4.346 Monetary value -.003 .000 .997 5.497 Age .014 .000 1.014 1.068 Gender -.065 .003 .937 1.038 Location ,000 ,010 1,000 1.002
The overall percentage of correctly classified predictions is equal to a hit rate of 86.0%. When a comparison is made between the developed model, a naïve model (p*p+(1-p)*(1-p); hit rate of 68.3 %) and a null-model (only the intercept is included; hit rate of 80.3%), it can be concluded that the estimated model is better in predicting the probability of a customer to churn than a random selection would have. Other studies using logistic regression to predict churn report hit rates varying from 40 to 70% (e.g. Yang and Chiu, 2006; Dasgupta et al, 2008; XIA and Jin, 2008). Compared to those hit rates, a hit rate of 86.0% is high.
34
reveals the proportion of customers that has the highest probability to churn as a specific group of customers. For instance, the customers from the top three deciles (30% of the customers) in the developed model will account for almost 75% of all churned customers from the total customer base (compared with 30% churned customers when randomly select 30% of all customers).
Figure 3: Cumulative lift curve with a Binary logit model
To provide an impression of the predictive usefulness of the model, predictive validity is determined. The dataset is divided into an estimation sample (which contains 95% random sampled customers) and a validation sample (which contains 5% random sampled customers). As presented in figure 3, the validation sample shows a similar cumulative lift as the estimation sample (Binary logit model) does. The validation sample has a hit rate of 84.8% and a Top-decile lift of 2.98, which are comparable to the estimation sample (with a hit rate of 86% and a Top-decile lift of 3.19).
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100 Per ce n tage ch u rn e rs Percentage of customers
35
5.2.4 Hypotheses
In order to find evidence about the importance of promotional behaviour on churn behaviour, the following hypotheses were proposed:
H1 Whether a customer receives a discount with his or her first purchase, has a positive effect
on a customer’s probability to churn.
H2 The extent to which the value of the total discount received as a percentage of gross sales
is higher, has a positive effect on a customer’s probability to churn.
H3 The extent to which a customer receives discounts more often, has a negative effect on a
customer’s probability to churn.
5.2.4.1 First purchase discount
As presented in table 3, the parameter estimate of First purchase discount differs significantly from zero (β = -0.078; Wald statistic = 15.631; p-value = 0.00). The odds ratio for First purchase discount is below 1 (Exp(-0.078) = 0.925), implying that if a customer receives a discount during his first purchase, this is associated with approximately 7.5% reduced odds of churning. The result shows that First purchase discount has a significant negative impact on the probability to churn instead of a positive impact as proposed in hypothesis 1.
36
Figure 4: Effect of First purchase discount on predicted probability to churn
5.2.4.2 Percentage of discount
As presented in table 3, the parameter estimate of Percentage of discount differs significantly from zero (β = 0.265; Wald statistic = 24.580; p-value = 0.00). Percentage of discount has an odds ratio of 1.303, which implies that an increase in Percentage of discount with 1 (=100%), results in an increase of approximately 30.3% in odds of churning. The result shows that Percentage of discount has a significant positive impact on the probability to churn: the extent to which the value of the total discount received as a percentage of gross sales is higher, has a positive effect on a customer’s probability to churn (with an increase in utility of β = 0.265). The findings of percentage of discount support hypothesis 2.
Figure 5 presents the effect of Percentage of discount on the predicted probability to churn visually. The figure is based on the means of the determinants stated in table 3, except from Percentage of discount (which changes along the x-axis) and First purchase discount and Gender (which are coded as zero’s). As can be seen in figure 5, the slope has an upward direction, indicating the positive effect of Percentage of discount on the predicted probability to churn. 0,0220 0,0225 0,0230 0,0235 0,0240 0,0245 0,0250 0,0255 No Yes Pr e d ic te d p ro b ab ili ty to c h u rn
37
Figure 5: Effect of Percentage of discount on predicted probability to churn
5.2.4.3 Number of discounts
As can be seen in table 3, the parameter estimate of Number of discounts differs significantly from zero (β = -0.049; Wald statistic = 37.726; p-value = 0.00). Furthermore, the odds ratio of Number of discounts is below 1 (Exp(-0.049) = 0.952), implying that an increase in Number of discounts with 1, results in a decrease of approximately 4.8% in odds of churning. The result shows that Number of discounts have a significant negative impact on the probability to churn: the extent to which a customer receives discounts more often, has a negative effect on the customer’s probability to churn (with a decrease in utility of β = -0.049). The findings of number of discounts support hypothesis 3.
38
Figure 6: Effect of Number of discounts on predicted probability to churn
A summary of the outcomes of the hypotheses is presented in table 4. Although promotions might reduce a customer’s probability to churn, the question rises whether more promotions will not have any negative effects on nett sales. As presented in Appendix A, a linear regression of Nett Sales is overall significant (F-stat = 40.210; p-value = 0.00) and shows significant results of promotional behaviour (respectively First purchase discount, Percentage of discount and Number of discounts: β = -4.749, value = 0.00; β = -13.639,
p-value = 0.00; β = 1.179, p-p-value = 0.00).
Table 4: Outcomes of hypotheses
Dependent variable Independent variable Hypothesis Proposed effect on churn Result
Customer churn First purchase discount H1 Positive Reject (but
significant negative) Percentage of discount H2 Positive Accept
39
5.2.4.4 Multicollinearity
Multicollinearity issues arise due the fact that independent variables correlate and it is therefore difficult to distinguish the effect of separate independent variables on the dependent variable (Leeflang, Wittink, Wedel and Naert, 2000). The degree of multicollinearity can be defined by the Variance Inflation Factor (VIF). When the VIF-score is below 10, there are no multicollinearity issues in the model (e.g. Neter, Wasserman and Kutner, 1989; Hair, Anderson, Tatham and Black, 1995). As presented in table 3, all VIF-scores are below 10. Frequency and Monetary value have the highest VIF-score. This makes seems intuitive, because a higher frequency might be related to a higher monetary value. However, both VIF-scores are too low to disturb the effect of the independent variables on the dependent variable. As mentioned before, the determinant Switching costs indirectly includes Frequency and Monetary value. However, as demonstrated in table 3, also Switching costs has a VIF-score below 10.
Since there are some unexpected signs in the parameter estimates (as discussed in chapter 5.2.3), it is evaluated whether these signs will change after leaving out those determinants with a high VIF-score. Except when leaving out Monetary value as well as Frequency, the results show that the signs of the parameter estimates do not change when leaving out Monetary value, Frequency and/or Switching costs. When Monetary value as well as Frequency are left out, the sign of Switching costs becomes negative (β = -0.068;
p-value = 0.00). However, both First purchase discount and Relationship length become
40
5.3 Cox regression model
5.3.1 Evaluation of overall model
From those customers who bought a product in the period from August 2011 till July 2012, approximately 84% also bought a product in the period from August 2012 till April 2014. In order to determine whether the determinants have a significant effect on future purchase behaviour (and whether a customer will survive), the Omnibus test of model coefficients is used (a comparison of the AIC between the Cox regression model without the determinants and with the determinants). This comparison results in a difference in AIC of approximately 15821 (= AIC without variables – AIC with variables: 2528972.984-2513151.631). Furthermore, the model is overall significant (Likelihood ratio/χ2 = 22769.519; p-value = 0.00) and has a significant change compared to the null model (Likelihood ratio/χ2 = 15821.353; p-value = 0.00).
5.3.2 Promotional behaviour
The Cox regression model assumes that the estimated beta’s of the determinants are constant over time and that the probability to purchase in the future can therefore be determined by the determinants at any point in time. In order to determine whether promotional behaviour influences the time when customers purchase, a Cox regression analyses is performed. As presented in table 5, First purchase discount, Percentage of discount and Number of discounts differ significantly from zero (Respectively First purchase discount, Percentage of discount and Number of discounts: β = -.004, Wald statistic = 12.762, p-value = 0.07; β = -.332, Wald statistic = 228.311, p-value = 0.24; β = 0.007, Wald statistic = 3856.139, p-value = 0.01).
41
Table 5: Results of the Cox regression model
Variable Beta P-value Exp(Beta) Mean
Promotional behaviour:
First purchase discount -.004 .007 .996
Percentage of discount -.332 .024 .717 .1227 Number of discounts .007 .001 1.007 2.62 Switching costs .002 .000 1.002 14.69 Customer dissatisfaction -.001 .002 .999 .42 Relationship length -.004 .000 .997 44.06 Recency .030 .001 1.031 8.48 Frequency .010 .000 1.010 26.94 Monetary value .000 .000 1.000 137.348 Age .002 .000 1.002 39.24 Gender -.053 .009 .948 Location .000 .000 1.000 8085,32
This suggests that a First purchase discount and an increase in Percentage of discount will cause a decrease in the probability to purchase in the future (i.e. survival function), and an increase in Number of discounts will cause an increase in the probability to purchase in the future (i.e. survival function). Figure 7 shows the survival curve of the model predicted time to purchase for the average customer (if all determinants are at their mean, see table 5). The probability to purchase rapidly decreases over the first ten months, but decreases only slightly from the 11th month. Therefore, the more time elapses, the lower the probability that a consumer will purchase a product.
Figure 7: Survival function of predicted probability to purchase
42
43
6. Discussion and conclusion
Churning customers have an impact on a company’s performance for two reasons: firstly, a decline in the number of customers is associated with revenue loss. Secondly, acquiring new customers is more expensive than keeping the existing customers (Reichheld and Sasser, 1990; Reinartz, Thomas and Kumar, 2005). In order to predict and reduce churning customers, previous studies tested and identified significant relationships between various determinants and a customer’s probability to churn (e.g. the length of a relationship, dissatisfaction, switching costs). Despite indications that there might be an effect between the relationship of a promotional behaviour and a customer’s probability to churn, this has not been tested before. Therefore, this study contributes to the existing literature on customer churn behaviour by studying the effect of promotional behaviour. The following research question was proposed: What is the effect of a promotional behaviour on a
customer’s probability to churn?
Several classification trees (CHAID, Exhaustive CHAID and CART) and a Binary logit model were used to identify important variables and to test the effect of promotional behaviour on churn. This study has been conducted in a non-contractual setting and the moment that a customer is considered as being churned is set to a fixed number of months. As a result, it is hard to define the exact moment in time when customers leave a company. However, it is known at what moment in time a customer purchases a product, which is the opposite of churning. Therefore, a Cox regression model was developed to predict if promotional behaviour affects a customer’s future purchase behaviour.
6.1 Implications
44
this study could be used as a starting point for new, elaborated research to promotional motives for customers to churn.
Companies could use the findings of this study to develop customised marketing strategies to encourage customers to stay. Those customers that could potentially decide to churn might be provided with individual discounts. Furthermore, unnecessary discounts could be prevented, saving companies potentially a considerable amount of money. As presented in figure 4-6, the average customer has a low probability to churn. By including variables for promotional behaviour, companies may be able to forecast a customer’s probability to churn more accurately. The following paragraphs will describe how the probability for customers to churn could be further reduced by changing a company’s promotional policy.
Bawa and Shoemaker (1987) report that price-sensitive consumers tend to be less loyal than consumers that are not price-sensitive. However, this study found that customers who receive a discount on their first purchase have a lower probability to churn than customers who do not receive a discount on their first purchase. Therefore, customers who purchase products on a discount during the acquisition phase are not automatically similar to price-sensitive consumers. In order to reduce future churn behaviour, companies should acquire new customers with products on a discount. However, the downside of discounts is that they reduce the company’s income, which should be compensated by a higher nett sales. As presented in table A4 (and chapter 5.2.4.3), the average nett sales for customers who receive a discount on their first purchase are lower than the nett sales for customers who did not receive a discount on their first purchase. For instance, a customer who receives a 10% discount on his first purchase, is associated with a resulting predicted nett sales of €117.32, while a customer who does not receive any discount on his first purchase is associated with a resulting predicted nett sales of €122.07 (ceteris paribus). This implies that managers will have to decide whether they want to decrease the probability of customers to churn or increase the nett sales. However, since it is more expensive on the long term to acquire new customers than keeping existing customers (Reinartz, Thomas and Kumar, 2005), companies should focus on reducing churn instead of increasing nett sales on the short term.
45
results on the effect of promotion during the entire relationship. As the percentage of discount on a customer’s total gross sales increases, a customer’s probability to churn increases as well. Consequently, in order to reduce the number of churning customers and build loyalty among customers, companies should not provide too high discounts. This result is very positive for managers, since lower discounts cost a company less money and reduces the probability of customers to churn at the same time.
Furthermore, a customer’s probability to churn is lower as the number of discounts increase. This implies that offering more promotions results in customers remaining loyal, which is in line with the findings of Gabler and Reynolds (2013). They reported a positive effect of the number of discounts on purchase intentions. Therefore, although discounts should not be too high, customers have a lower probability to churn when receiving discounts more often. In addition to this, a higher number of discounts will result in higher nett sales as well (table A4).
As can be seen in figure 7, the more time elapses, the lower the probability that a customer will purchase again from the same company. One might expect opposite results of promotional behaviour on the probability of purchase and the probability of churn. However, First purchase discount has a negative effect both in relation to churn as to repurchase (no opposing signs). This effect might be caused by the time that passes. If a customer purchases his first product with a discount, the probability to repurchase in the future will be lower than for the customer who did not receive a discount on his first purchase. The Percentage of discount and the Number of discounts do show opposing signs.
6.2 Limitations and further research
46
Promotional behaviour can be explained using three different variables, which are all based on discounts: First purchase discount, Percentage of discount and Number of discounts. However, there may be more variables related to promotions which have an effect on customer churn. For instance, it is not clear what the effect of advertisement on customer churn is. Will more advertising cause less or more churn behaviour, and what is the effect of advertisement relating to competitors?
Although many determinants identified in prior studies were included in the model, the model explains whether a customer is considered as being churned ‘only’ up to 43% better than a null model would do. This implies that there are still some omitted variables that could increase the explaining power of the overall model. For instance, customer interaction and negative word of mouth might have an influence, which could stimulate customers to switch to other companies. In addition, Svendsen and Prebensen (2013) report that a positive brand image reduces customer churn probabilities. However, competition is not included in the model.
