EMPIRICAL GENERALIZATIONS ON THE
DRIVERS OF CUSTOMER CHURN
Master Thesis Business Administration, Specialization Marketing Management
EMPIRICAL GENERALIZATIONS ON THE
DRIVERS OF CUSTOMER CHURN
Author: Boudewijn Siekman
Student Number: 1278827
Supervisor: dr. J.E. Wieringa Second-assessor: drs. H. Risselada
Date: August 19, 2011
Faculty: Economics and Business
Course: Master of Sciences Business Administration
Profile: Marketing Management
Management Summary
Numerous studies have been published on the relationship between drivers of churn and resulting switching behavior. Empirical research into churn has given insight in the outcomes of relationship marketing, across products, markets, and regions. This study presents generalizations on the effectiveness of churn drivers. A meta-analysis has been performed with churn driver elasticities as the dependent variable. Churn driver elasticities are defined as the ratio to which churn changes by the change in churn driver variables. The meta-analytic regression is executed on 223 churn driver elasticities that were obtained from 23 empirical studies on churn. The independent variables are sixteen moderator effects that are obtained from the meta-data of the research base. They can be classified in five categories: driver characteristics, relationship characteristics, market characteristics, research methodology, and publication characteristics. Hierarchical linear models are used for estimation of the model using restricted maximum likelihood estimation (REML).
The results of this study indicate that churn is inelastic and differs among churn driver categories. The overall mean value of the observed churn elasticities is 0.226, with a standard deviation of 0.307. Attitude perceptions have a bigger impact on churn elasticities than behavioral antecedents. Of the churn driver variables, affective commitment and relationship characteristics have the greatest influence on churn. Moreover, situational factors, such as demographics, markets and industry are major influencers on churn elasticities. Interestingly, regions do not show a significant influence on the dependent variable.
Preface
In front of you lies my final work as a student of the University of Groningen. Handing in this thesis for the Master Business Administration, specialization marketing closes a period that was both hard work and fun at the same time. Although my studying years in Groningen lay behind me for a while now, this thesis closes the book finally. I want to thank some people that made this possible.
First of all, I would like to thank my thesis supervisor Jaap Wieringa. Without his help and guidance on the subject of churn, my efforts would not have been fruitful. His suggestions, comments, and patience have resulted in this document, which has taken longer to finish than both of us had expected. Moreover, I would like to thank my co-assessor Hans Risselada for his help in the final stages of this report.
Table of Contents
1. Introduction ... 9
1.1 Aim of this Research ... 9
1.2 Research Objective ... 10
1.3 Research Outline ... 11
1.4 Thesis Outline ... 12
2. Theoretical Underpinnings and Conceptual Model ... 13
2.1 An Introduction to Churn Theory ... 13
2.2. Customer Asset Management ... 15
2.3 Towards a General Model ... 17
3. Drivers of Churn ... 20
3.1 Complex interrelations ... 20
3.2 Behavioral Antecedents ... 22
3.3 Attitude Perceptions ... 24
3.4 Demographic Characteristics and Other Drivers ... 26
3.5 Conceptual Model ... 27
4. Data and Research Design ... 29
4.1 Inclusion Criteria ... 29 4.2 Literature Search ... 30 4.3 Dependent Variable ... 32 4.4 Research Base ... 34 4.5 Coding ... 36 4.6 Moderating Variables ... 37 4.7 Analysis ... 42 5. Results ... 44 5.1 Driver Characteristics ... 45 5.2 Relationship Characteristics ... 48 5.3 Market Characteristics ... 48 5.4 Research Methodology ... 48 5.5 Publication Characteristics ... 50 6. Discussion ... 51 6.1 Management Implications ... 53
6.2 Limitations and Future Research ... 54
1. Introduction
Customer churn is a reality all businesses face. That's why firms need to ask themselves: Will my customers return? What are the reasons for their behavior? How can we influence this behavior? Improved customer retention is therefore one of the key outcomes of relationship marketing.
Churn, and its counterpart retention, forms the central topic of this thesis. The dimensions and antecedents of churn and actions that firms can take to improve the chance that customers stay are discussed. Using a meta-analytic approach, the empirical results of churn research are synthesized in order to make generalizations on consumer churn.
1.1 Aim of this Research
What is the role of churn in the current developments in the field of marketing? Extensive research has been done in the field of consumer churn (Keaveney, 1995; Capraro et. al., 2003; Lemmons & Croux, 2006). Advanced statistical models are used to predict churn, using techniques such as logistic regression models and decision trees, for example in the churn tournament results of Neslin et al. (2006). Empirical research has also been performed to test churn models in practical market situations, such as switching behavior in liberalized markets (Wieringa and Verhoef, 2007).
Furthermore, several marketing models contain a churn component. The customer valuation model as used by Gupta, Lehman and Stuart (2004), uses retention as a specific variable to calculate CLV. They even find that the elasticity of churn in relation to lifetime value is up to five times higher than the discount rate for more mature businesses. Also conceptual models, such as the CUSAMS framework (Bolton et al., 2004), contain a churn (length) component. Moreover, marketing research has a focus on developing techniques for improving accountability and ROI (MSI 2008 - 2010). Finding ways to quantify the variations of the customer base is very much in line with this research objective.
advanced statistical models (Neslin et al., 2006; Lemmens and Croux, 2006) and papers that have the aim to test hypotheses on the relation between drivers and churn empirically (Gustafsson et al., 2005; Bolton, 1998; Verhoef, 2003). Two of the directions for future research by Bolton et al. (2004) are finding the key drivers for customer behavior and perform meta-analysis on empirical studies. Although explorative analysis that poses research questions for a generalized churn model is published (Verhoef, Van Doorn, Dorotic, 2007), these questions have not yet been answered.
This research aims to answer this call for more generalized conclusions in the field of churn research. Since a large number of empirical studies that analyze churn have been published, the similarities and differences between these studies should be analyzed. A meta-analytic approach has been chosen to find the building blocks for a general churn model.
1.2 Research Objective
To analyze whether general findings can be identified in the dimensions of churn, the theories of churn are explored and statistical analysis is performed using parameters and meta-data from articles on empirical retention research. This meta-analysis will be used to see whether building blocks for a general model can be found. From these delineations the main research objective is defined:
Advancing the general understanding of the drivers that influence churn by conducting a meta-analysis of empirical churn research.
This is quite a broad objective, as is required by the nature of this type of research. Several sub-questions can be derived from the research objective:
What literature and/or theories exist on the subject of churn? What empirical research has been done in the past few years? What conceptual model should be used to evaluate this research? Which churn driver variables are researched?
Which moderating variables can be identified?
What type of statistical analysis should be performed?
To date, no meta-analytic research on consumer churn has been performed. Neslin et al. (2006) do synthesize churn with data obtained from a tournament, which they refer to as 'meta-analysis'. However that research is aimed at improving model building. Therefore this research will be a valuable contribution to the current literature on the subject.
This paper is aimed at gaining insight in the causal relationship between churn drivers and churn. This study is also useful for managers who want to explore churn management as a marketing strategy, as it offers a broad overview on the subject and aims to give generalizations on the relationship between marketing drivers and churn.
1.3 Research Outline
Getting more general insight in churn requires that the meta-analysis is embedded in the current literature and the right research design is chosen. See figure 1.1 for the modeling flow. To perform the meta-analysis, a number of steps are taken. The data is obtained by searching for publications that quantitatively research the relationship between marketing variables and churn intention. Following the approach from Bijmolt and Pieters (2001) and Kremer et al. (2008), the published parameters are transformed to elasticities. This makes the parameters comparable between studies.
The metadata of the articles are coded. After some necessary adjustments, the final dataset is created (Lipsey and Wilson, 2001). As suggested by Bijmolt et al. (2005) categorical principal components analysis (CATPCA) is performed to check for correlations between the variables. After which a linear model with restricted maximum likelihood estimation (REML) is estimated to find what the relationship is between the metadata and the elasticities. The results of the analysis will be used to answer the research question.
Figure 1.1: Modeling Flow
Publication Search calculations Elasticity Coding of Metadata confounds Check for (CATPCA)
1.4 Thesis Outline
2. Theoretical Underpinnings and Conceptual Model
Retention can be defined as “the chance that the account will remain with the vendor for the
next purchase, provided that the customer has bought from that vendor in each previous case.” (Bund Jackson, 1985 p. 18). Retention, and its reciprocal churn1, has been researched quite extensively in the recent past. Retention goes to the heart of the current customer centric marketing paradigm (Vargo and Lush, 2004): Do customers return? And if so, why? For if one wants to build long-term relationships with their customers, these customers need to repurchase, or at least have that intent (Evanschitzky et al., 2006). In this chapter, the relevant theories related to churn are discussed.
2.1 An Introduction to Churn Theory
As early as 1960, Theodore Levitt asked the question to organizations: “What business are
you really in?” He argued that: “...the organization must learn to think of itself not as producing goods or services but as buying customers, as doing the things that will make people want to do business with it.” In other words, why would customers buy in the first
place and come back later. Together with scholars as Kotler (1967) and McKitterick (1957) these were the first steps toward a more relationship-orientated view of marketing.
In the 1980s, the term relationship marketing was first introduced. The American Marketing Association defines relationship marketing as: “Marketing with the conscious aim to develop
and manage long-term and/or trusting relationships with customers, distributors, suppliers, or other parties in the marketing environment.” From this paradigm, returning customers got
more attention (Bund Jackson, 1985). Bund Jackson split customers into two groups: always-a-share, and lost-for-good. The always-a-share customers can easily switch between suppliers and diversify their purchases. Migration models are used to describe the always-a-share behavior. The time since last purchase (recency) is used to predict the chance of a repeat purchase. This means that a purchase period can be skipped; only the probability of return can change. Lost-for-good customers are bound to one supplier. No other vendors are purchased from. When customers do not return the next period they are considered lost for good. If they
1
come back later, they are considered as newly acquired. The main reason why customers are classified in the lost-for-good group is the existence of (perceived) switching costs. To model lost-for-good situations churn models are used. These models use non-time variables such as historical purchase data to predict the probability of return or retention rate.
Others also explored advantages of churn in marketing. (Rosenberg and Czepiel, 1983; Fornell and Wernerfelt, 1987). A series of articles published mostly by Reichheld (Reichheld and Sasser 1990; Reichheld 1993; Jones and Sasser 1995; and Reichheld 1996) and a book (Reichheld 1996a), gave arguments for why churn management created more value from customers. Retaining customers could be profitable for six reasons:
1. Lower acquisition cost: Returning customers do not require the high cost of acquisition.
2. More revenue: long-term customers provide more revenue.
3. Lower handling cost: Mutual learning practices lower need for employee attention. 4. Referrals: Long-term customers give more referrals to potential customers.
5. Pricing: Loyal customers will accept higher prices
6. Defection: Customers who stay longer have a lower chance of defection. This increases the value of future earnings.
Reichheld argued that long-term customers would have a higher lifetime value. Therefore the proportion of long-term customers in the customer base of firms should be increased. This is done via marketing efforts that raise satisfaction, increase purchases, and prevent churn.
Not only the applicability of single-focus churn strategies, but also their interrelatedness were questioned. Reinartz and Kumar (2002) showed, for example, that long-term customers were not cheaper to serve. A lot of the statements made by Reichheld were case-based. Although memorable, the anecdotal nature of the cases does not make them generally applicable. In practice, it seems, churn management has less advantages than posed by Reichheld.
East, Hammond and Gendall (2006) give a full overview of the reasons why churn management may have been approached overenthusiastically. Their main conclusions are that the value of long-term customers varies greatly and that the benefits of the long-term customers are exaggerated. They propose an approach to increased consumer value where churn management is one of the many roads to be taken.
Churn management has to be viewed within a broader, interconnected view of marketing. This approach takes into account a more embedded role of churn. However, the assumption is often made that increased customer (quality) satisfaction is the main driver for lower churn and higher value. However, the relationship between satisfaction and customer lifetime value can be weak (Reinartz and Kumar, 2000).
2.2. Customer Asset Management
Although the CUSAMS framework does offer an intricate view of the relationship between organizations and customers, this paper focuses on the drivers that influence consumers in their decision to stay or defect.
Figure 2.2: The CUSAMS framework
2.3 Towards a General Model
Verhoef, Van Doorn, and Dorotic (2007) state four questions to come to a general model of churn. These questions are:
1. Can we develop a general model explaining customer churn? Which models predict customer churn best?
2. How does the effect of the determinants of customer churn differ between industries and market environments?
3. How does the effect of the determinants of customer churn differ between countries? What is the role of culture and a country’s economic characteristics?
4. Which factors determine customer churn under changing market conditions, such as market liberalization? What are the impact of market conditions on customer churn and its determinants?
Paul et al. (2009) give building blocks for such a model. Via laddering interviews they obtained a comprehensive list of drivers of repeat purchases (see table 2.1). Their qualitative research provided a framework in which consumers were asked what drove their repeat purchases. Using the means-end theory by Gutman (1982) they arrive at a framework that identifies three interlinked categories of twelve first-order and fifty-one second order drivers: Motivational values, relationship driving benefits, and service relationship attributes. They find that on the service relationship side, having a good product and delivering good service is a ‘conditio sine qua non’ for establishing long-term relationships. Their research also concludes that repeat purchases are driven mostly by motivational values. Motivational values are mostly ignored by current research, most likely due to the fact that they are hard to measure quantitatively.
Table 2.1: First order Drivers of the Repeat-Purchase driver Model
Driver Category: First-order Drivers Service relationship attributes Service product
Service delivery Service environment Service location
Relationship-driving Benefits Relationship characteristics Company characteristics Functional benefits Psychological benefits Social benefits
Motivational Values Individual motivational values Collective motivational values
Mixed motivational values
Figure 2.4 Framework for Customer Loyalty
Bansal et al. (2005) also proposed a framework for switching behavior. This framework uses a migration model to analyze switching behavior. Push and pull variables are the main drivers for migrations, while intervening (mooring) variables either facilitate or inhibit the migration. They find that the emphasis in current research focuses to much on push effects (such as satisfaction, quality perception), while the influence of pull effects such as competing services and the moderating role of the mooring variables such as switching costs are undervalued in current research.
3. Drivers of Churn
In this chapter, the drivers of churn that are used in the empirical marketing literature are discussed. These drivers will be explored qualitatively to lay the foundation for the meta-analytic research. After a brief introduction of empirical churn research, the empirical research is split in research streams and discussed. Several characteristics that are relevant for the meta-analysis are explored subsequently. The chapter concludes with the conceptual model that will be used for the research.
3.1 Complex interrelations
One of the early publications for which broad data was collected on why customers switch was an exploratory research by Keaveney (1995). A critical incidents analysis provided a model that identified reasons for consumers to switch. The analysis showed that switching is a complicated process, 55 percent of switching resulted from multi-factor incidents. Bolton (1998) also concludes this. The paper states on the relation between satisfaction and churn: “Managers and researchers may have underestimated the importance of the link between customer satisfaction and churn because the relationship between satisfaction and duration times is very complex and difficult to detect without using advanced statistical techniques.”
Table 3.1 shows a list of empirical papers that have been considered for the meta-analysis and the types of determinants that are used in these publications. The table shows that most research focuses on perception drivers to explain churn. Only two publications solely use behavioral antecedents two explain churn.
Table 3.1: Empirical Publications on churn and the types of churn drivers used.
Author Behavioral Perception Demographic
Abdul-Muhmin (2005) X
Anderson (1994) X
Anton et al. (2007a) X X
Anton et al. (2007) X X
Athanassopoulos (2000) X X X
Bansal and Taylor (2002) X
Bansal et al (2005) X
Bansal et al. (2004) X
Bhattacharya (1998) X
Blackwell (1999) X
Bolton, Kannan, Bramlett (2000) X X
Bolton, Lemon, Bramlett (2006) X
Burnham et al. (2003) X
Capraro, Broniarczyk, Srivistava (2003) X X
Gounaris (2005) X
Gustaffson, Johnson, Roos (2005) X X
Hansen et al (2003) X Haughton et al. (2006) X Hellier et al (2003) X Jamal Bucklin (2006) X X Jones et al (2000) X X Jones et al. (2003) X Jones et al. (2007) X
Lemon, White, Winer (2002) X
Liang et al. (2009) X
Mittal et al. (1994) X
Mittal, Kamakura (2001) X X
Money (2004) X
Ranaweera & Prabhu (2003) X
Risselada, Verhoef, Bijmolt (2010) X X
Sanbandam and Lord (1995) X
Seiders et al. (2005) X X X
Spreng et al. (1995) X
Vazquez-Casielles et al. (2009) X X X
Verhoef (2003) X X
Verhoef, Langerak, Donkers (2007) X
Wieringa, Verhoef (2007) X X
Zeelenberg en Pieters (2004) X
3.2 Behavioral Antecedents
Behavioral antecedents are variables that show the relationship between actual behavior and churn. The relationship between actual behavior and interactions between the customer and the firm or brand and churn are often very clear: Objective measures give customers a rational way to evaluate the decision to return. The research by Keaveney (1995), for example, shows the relationship between a critical incident and the resulting switching behavior.
Due to the data-driven nature of Behavioral Antecedents the variables consist of actual data from databases or surveys that are mostly continuous or binomial. The industries that are researched are data driven and require regular repeat purchases, such as financial markets or telecommunications. The statistical techniques used are also quite advanced. Logistic regression is used more than linear regression techniques.
Critical Incidents
Critical incidents are actual 'grounded events' that might cause customers to switch. Keaveney (1995) defines them as: "...any event, combination of events, of series of events between the customer and one or more service firms that caused the customer to switch service providers." Variables that are used to measure critical incidents are: The number of times a technician has to be used, and how long it takes to solve the problem (Bolton, Lemon and Bramlett, 2006) how often a phone call gets disconnected (Seo et al., 2007; Ahn et al., 2006), or how often someone has become angry with a company (Anton et al., 2007).
Relationship Age
The previous experience of customers with a firm can influence whether they switch or not. Customers will update their expectations and future value of a product or service with the experience they had with this firm in the past, as described in the theory of belief updating by Hogarth and Einhorn (1992). Customers who have a longer relationship with a provider will know what they can expect from their service provider. This reduces insecurity, which the customer would have with another provider. Relationship age is shown to have a negative influence on churn (Bolton, 1998) and lifetime value (Reinartz and Kumar 2000, 2003). Moreover, as Verhoef (2003) states: "Past customer behavioral variables, can also be indicators of past behavioral loyalty, which often translates into future loyalty". Therefore a longer relationship is expected to lead to lower churn.
Relationship Marketing Instruments
The nature of the customer centric paradigm is to bind customers to engage in a mutually valuable relationship. To do so, firms use Relationship Marketing Instruments (RMIs), such as loyalty schemes and direct mailings (Hart et al., 1999). These programs are aimed at improving the attitude of the customer toward the company and incentivizing them to maintain a long-term relationship with the company. However, there is evidence that only customers who already perceive the service level to be valuable will use these programs. These programs need to be seen as complementary within a Relationship Marketing Strategy (Bolton, Kannan, Bramlett, 2000). Since customers who join a loyalty program are expected to be long-term customers who are more satisfied with the firm, relationship-marketing instruments are expected to have a negative relationship with churn.
Other Behavioral Antecedents
3.3 Attitude Perceptions
The choice of a product or brand is seldom a purely rational one. Even though markets may be considered rational, rationality is seldom seen with individuals. A large part of the quantitative literature researched includes personal beliefs and attitudes toward a product or brand: Attitude Perceptions (Bolton 1998, Capraro et al. 2003, Mittal & Kamakura 2001, Seiders et. al 2005). Marketing researchers ask consumers how they feel about a product, brand, or firm, and use statistics to find out whether this attitude has any influence on their behavior, in this case churning.
This type of research is about perceptions, and quantifying impact is difficult. Therefore variables used in surveys and questionnaires are mostly ordinal or categorical. The aim is not only to find a quantitative relationship, but also to explore the complex interrelationships. To do this, not only (logistic) regression is used (Wieringa and Verhoef, 2007), but also structural equation modeling (Bansal et al., 2005). The industries researched are often less data-driven, such as automobile, retail, which result in non-contractual relationships. However, markets as telecommunications, energy, and banking are often researched as well.
Commitment
Commitment can be defined as a force that binds an individual to a course of action of relevance to one or more targets (Meyer and Herscovitch, 2001). Three types of commitment are defined: Affective, Calculative and Normative (Meyer and Allen, 1997). Affective commitment is a commitment based on desire. Calculative commitment is based on the need to maintain the relationship, for example due to low cost of a product. Normative commitment is based on the fact that a consumer feels obliged to stay; it is the right thing to do (Bansal et al., 2004).
The researched quantitative literature only used affective and calculative variables as drivers of churn. Jones et al. (2007) give an elaborate overview of past research about affective and calculative commitment. Since commitment is a binding force, it is expected that both affective and calculative commitment have a negative relation with churn.
Involvement
importance of the product to the consumer means that high involvement relates to personal relevance where low involvement means that the consumer is indifferent (Mittal, 1995).
Seiders et al. (2005) relate involvement to churn (intentions), where the expected behavior is not clear. More involved customers are expected to be more likely to return, since they made a choice based on the relevance, and resources spent to come to the decision. However, since high-involved consumers have more information about alternatives, and possibly higher expectations repurchase intentions maybe negatively influenced. Seiders et al. also find evidence for this, where certain groups of low-involved, yet satisfied customers have high repurchase intentions.
Satisfaction
Satisfaction can be defined as: "The overall evaluation based on the total purchase and consumption experience with a good or service over time" (Anderson et al., 1994). In short: are the customers' expectations of a product or service met? Satisfaction has been positively associated with self-reported loyalty in the meta-analysis of Szymanski and Henard (2001). Verhoef (2003) also gives an overview of publications that use satisfaction to model behavioral loyalty. This overview shows that satisfied customers are expected to have a negative influence on churn.
Relationship Quality
Relationship quality is a term that was used by Verhoef and Wieringa (2007) in a construct that combined a number of variables associated with the quality of the relationship between a customer and the firm such as: Trust, service quality, product quality and word-of-mouth (WOM). These variables are considered as antecedents of satisfaction and commitment in some studies (Verhoef, 2003; Szymanski and Henard, 2001). However, like satisfaction and commitment, relationship quality variables are also often included as a direct influence on churn (Foscht et al., 2009; Anton et. al., 2007; Bansal et al., 2005). Relationship quality is expected to have a negative impact on churn.
Switching Risk
with switching risk are: Financial cost associated with switching (Hu et al., 2006), general attitude toward switching (Bansal et al., 2005), and perceived switching costs (Wieringa and Verhoef, 2007). Switching risk is expected to have a negative relationship with churn.
Other Attitude Perception Drivers
Finally, there are a number of perception drivers that are used as drivers for churn in the researched publications. These were variables related to the knowledge of the product (Capraro et al., 2003), repurchase intentions (Bolton et al., 2000), convenience (Seiders et al., 2005), and sophistication (Maicas, Polo, Seze, 2009). These are expected to have a negative association with churn.
3.4 Demographic Characteristics and Other Drivers
Quantitative marketing analysis uses demographic characteristics of subjects as moderating variables in marketing research to analyze whether subgroups may exist within the sample or population or to research what kind of characteristics exist within a consumer segment. Variables that are commonly used are age, gender, income, marital status, and income. For example, a demographic can state what the preferences are of an unmarried 25-year-old male, with a middle-income. The expected relationship between churn and demographics cannot be stated, since demographics are of a descriptive nature and, and as a general rule, are expected not have a causal relationship to churn. Notwithstanding that for specific products, markets, or categories demographics can have influence on churning behavior.
3.5 Conceptual Model
To analyze what the influence of meta-data is on churn drivers, a meta-analysis is performed. The above paragraphs and chapter two have provided an overview of relevant churn theories and drivers (see table 3.2 for overview). This paragraph will present how meta-data may influence these churn drivers. This is the last building block for the conceptual model after which it will be presented.
Table 3.2: Churn drivers and expected relation to churn
Type of Effect Churn Driver Expected*
Behavioral Antecedents Critical incident +
Relationship Age -
RMI -
Other Behavioral -
Attitude Perceptions Affective Commitment -
Calculative Commitment - Involvement +/- Relationship Quality - Satisfaction - Switching Risk - Other Perception -
Demographics Demographics n/a
Other Drivers Other drivers n/a
* Expected relationship in relation to churn
Figure 3.1: Conceptual Model
The effect size statistics derived from the churn driver variables in this analysis are influenced by metadata which can be classified in five categories: driver characteristics, relationship characteristics, market characteristics, research methodology, and publication characteristics. These classifications are chosen as similar meta-analytic research (Tellis, 1988; Bijmolt et al., 2005; Kremer et al., 2008) uses similar categories.
The conceptual model uses this metadata and uses them as moderating variables that influence the churn driver effect size statistic as shown in figure 3.1. The next chapter that discusses data and methodology will also discuss the calculation of the effect size statistic, moderating variables and their expected relation in greater detail.
4. Data and Research Design
In order to find all published and unpublished empirical studies on churn, a thorough search strategy has been executed. The selection criteria and search method are stated first, after which the included studies are discussed. The dataset coding and further adjustments are reviewed subsequently. Finally, the further steps in the research design are described: the dependent and moderating variables, and the statistical analysis techniques chosen.
4.1 Inclusion Criteria
The objective of this research is to give a broad insight in empirical churn research. We have chosen for an inclusive approach that uses studies from a broad number of research methods, data types, and driver variables. The author is aware that this can be cause for an argument with regards to the meta-analytic validity and robustness of the research. However, the need for general insights, which overlap research methods, industries and different demographic groups, is considered to outweigh this argument. Moreover, there is also a convenience argument. Although quite a lot of research has been published on churn, the fragmented nature of conceptual models, research designs and data categorization makes it currently impossible to adhere to requests for a restrictive approach.
For a study to be included in the meta-analysis, it must meet several criteria. Lipsey and Wilson (2001, p. 16-23) give a detailed discussion on the eligibility criteria for a meta-analysis. These criteria are used for this study and applied to empirical churn studies below.
Eligible studies need to study churn between consumers or firms and a formal organization that delivers a product or service. This can either be a profit-seeking or non-profit organization. Respondents can be of any age, demographic or region.
(WOM) are not included in the study. The independent driver measures (behavioral, perception, or demographics) can also be actual or self-reported and also be constructs or single variables. The relationships must be direct. Interaction and indirect measures are not included in the research. It must be noted, that due to different definitions of the measures included in the studies, validity threats may arise. Again, this issue is recognized, however due to the objectives of the research, the choice is made include a broad set of measures.
As the research design of this study uses elasticities as the effect size statistic. More details on the choice for the effect size statistic and how they are calculated are given in paragraph 4.3. Included studies need to use research designs that result in measures that can be calculated to elasticities. Therefore studies that use advanced statistical measures such as decision trees (Neslin et al. 2006) are not included.
The time frame of this search is publications from 1985 onward. Barbara Bund Jackson first published the definition of retention in that year. Therefore only articles published after that date are included.
Studies may be published or unpublished. A publication can be any type of journal, refereed, non-refereed, working papers, dissertations, and master theses. This is in line with the need for a broad scope of the research.
4.2 Literature Search
The aim of the data collection is to find all published and unpublished articles that perform empirical studies on churn. We have chosen for a multi-faceted strategy, which includes combinations of cross-reference search and academic search engines.
Consequently we performed an online bibliographic search. We used synonyms of churn and retention as main keywords: ‘churn’, ‘retention’, 'switching', 'defection', 'attrition', 'repeat-purchase', 'repurchase likelihood', and 'repatronage'. However, using only a single keyword in a query gave unsatisfying search results, as most definitions have a different use in other types of academic research. Therefore they were combined with other keywords so the queries would be more marketing-related. The words used were: 'consumer', 'customer', 'intention', 'behavior', ‘prediction’, ‘marketing’, and 'brand'. The keywords were also combined with relevant marketing journals such as: Journal of Marketing, Journal of Marketing Research, Marketing Science, Management Science, and Journal of Services Marketing.
The main search engines used were Google Scholar and PurpleSearch. On the articles found with Google Scholar, the ‘cited by’ function was used to find out whether relevant follow-up research had been done. One of the advantages of Google Scholar is that it also indexes hard to find working papers. The PurpleSearch engine from the Library of Groningen is an aggregating search engine that enables simultaneous search in multiple academic databases. We have used Business Source Premier, Academic Source Premier, PiCarta and the University of Groningen Library Catalogue. The main advantage of PurpleSearch is that results are ordered by year of publication, instead of relevance due to academic citations. This is more in line with the aim of the search to find and include as many studies as possible, as you can search through the databases chronologically. Furthermore, search queries were executed in the databases of Ebscohost and Science Direct.
Finally, the author examines the personal websites of a number of authors that published multiple articles in this field to find their list of publications, mainly to find unpublished working papers. These are the authors that have been investigated: R. Bolton, H.S. Bansal, and M. Jones, S. Keaveney, S. Neslin, Z. Jamal, and V. Kumar.
4.3 Dependent Variable
As stated in paragraph 3.6, meta-analysis can be viewed as derivative research, where it's assessed to what respect the circumstances, or meta-data (such as relationship type, market characteristics and research design), in quantitative studies influence the outcome of the research. However, most studies are not directly comparable because of differences in research designs. Meta-analysis relies on statistical standardization to make the results comparable. This is achieved by calculating a dependent variable that is comparable across studies, called the effect size statistic Lipsey and Wilson (2001).
For this meta-analysis on churn drivers, several effect size statistics were considered. Lipsey and Wilson (2001), provide rationale for deciding which effect size statistic to use. In an 'ideal-world-situation' the effect size statistic used would be the Fischer z-transformed correlation. Unfortunately, most studies do not publish sufficient data to calculate this measure. Therefore, churn driver elasticities have been chosen as the effect size statistic for this meta-analysis.
Elasticity is the ratio of the percentage change in one variable to the percentage change in another variable. In this research, churn driver elasticities are defined as the ratio to which churn changes by the change in churn driver variables. Elasticities may not hold as much information on the relationship between drivers and churn as correlations. However, they are insensitive to the unit of measurement, and can therefore be used as a measure to compare studies (Bijmolt and Pieters, 2001).
Churn driver elasticity measures what the influence is of a churn driver variable on churn. Using these as dependent variables in statistical analysis gives shows the level to which these variables influence churn across publications. To validate the influence of the churn drivers, moderating variables are added to analyze the extent to which study-specific characteristics of churn research influences elasticities and therefore the level to which they can be generalized.
Elasticity Calculations
recoded to elasticities: Ordinary least squares regression (OLS), structural equations modeling (SEM), logistic regression (logit), and binary response models (probit).
The elasticities from OLS and SEM regression parameters were calculated using the elasticity formulas provided by Gemmil et al. (2007), and are stated in table 4.2. Structural models posed some issue, since they are dynamic models where independent variables can have both a direct and indirect influence on churn. To avoid problems with multicollinearity and simplify the calculations, the indirect variables of structural models have been omitted from the analysis.
Table 4.2: Elasticity calculations for OLS models
Regression Specification: Statistical Model Elasticity Formula
Linear y = yo + y1x + e y1 (!/!)
Log-log ln(y) = yo + y1 ln(x) + e y1
Linear-log y = yo + y1 ln(x) + e y1(1/ !)
Log-linear ln(y) = yo + y1 x + e y1 !
The elasticity formulas stated above are not applicable to logit and probit models. These models do not estimate a straight line as the least squares regression, but an s-shaped curve, or signoid function that forces the dependent variable to be between zero and one. Especially for churn research, this can be useful, since the outcome of churn marketing is binary: customers either churn or retain. However, since the shape of the curve is not straight, the value of the elasticity will always be dependent on the values of the variables. These elasticities are called quasi-elasticities. It is usual to calculate the quasi-elasticities with the mean value of x. To calculate the specific values of the elasticity for logit models, the following formulas are used (Franses and Paap, 2001, page 58):
Quasi elasticity: ! ! = !!! Pr (!)(1 − Pr ! )
Where:
Probit models use binomial data as input for the variables the calculations are different. To calculate the quasi-elasticities, the probability density function of the normal distribution is used (Franses and Paap, 2001, page 118). Similar to the logit functions, the average values of the variables are used to calculate the elasticities.
Quasi-elasticity: ! ! = Φ !′!! !!
Where:
And: !′!! = (1 !) !!!
!
4.4 Research Base
The search resulted in a list of 287 studies that have been actively considered for the meta-analysis2. Unfortunately, upon further investigation, a large number of studies that seemed eligible at first did not meet the inclusion criteria. The publications that did meet these criteria resulted in a list of 58 articles that report 447 measures that either had behavioral antecedents, attitude perceptions, or demographics as an independent variable and a churn metric as independent variable.
Quite a few articles do not publish sufficient data to calculate churn driver elasticities. The number of elasticities resulting from the eligible articles is 185 parameters from 24 publications.
Publication Bias
Although efforts have been made to include all papers that quantitatively research churn, it has to be noted that the parameters found in the study may be an overstatement of reality. Studies that present negative results (i.e. no variation in the outcome) may remain unpublished and can be difficult to find. This publication bias is the name for what occurs whenever the research that appears in the published literature is systematically unrepresentative of the population of completed studies (Rothstein et al., 2005). This study aimed to find unpublished publications as well, however, none of the found articles passed the inclusion criteria. To check for the publication bias, a funnel plot has been made, see figure 4.1. This scatter graph plots the number of observations against the effect size statistic.
2 There was a big overlap in the results of search methods. The gross list of considered studies contained over
five hundred papers.
Φ(!̅′!!) = (1/!2Π!)exp (−1 2(
Figure 4.1: Funnel plot with outliers
Figure 4.2: Funnel plot of final dataset
The final list of elasticities is given in table 4.1. It consists of 24 publications that report 183 drivers of churn. Since demographics and other driver variables of the models are also taken into account in the statistical analysis, the full sample size is 223.
Table 4.1: Articles included in meta-analysis
Author Year Count and Average Elasticity reported
Behavioral Perception Demographic Other drivers
Abdul-Muhmin 2005 2 -0.060 Ahn et al. 2006 10 -0.001 4 -0.096 Anderson 1994 1 0.573 Anton et al. 2007 2 0.106 22 0.067 Athanassopoulos 2000 9 0.055 Bansal et al 2005 12 -0.296
Bolton, Kannan, Bramlett 2000 2 -0.011 7 -0.006
Bolton, Lemon, Bramlett 2006 12 0.014
3 -0.033 Burnham et al. 2003 2 0.320 Capraro, et al. 2003 10 -0.087 Foscht et al. 2009 6 0.011 6 0.841 Gounaris 2005 2 0.317 Hu et al. 2006 5 -0.438
Ittner and Larcker 1998 1 0.002 2 0.071
Jones et al. 2007 8 0.158
Lopez, et al 2006 4 -0.005
8 0.061 4 -0.012
Maicas, Polo, Seze 2009 3 0.014 1 0.024 1 0.011 2 0.072
Seiders et al. 2005 2 0.000 3 0.264 1 -0.136 1 0.000 Seo et al. 2007 4 -0.022 4 -0.069 4 -0.122 Vazquez-Casielles et al. 2009 2 0.015 5 -0.123 1 0.023 Verhoef 2003 4 0.007 6 -0.007
Verhoef, Langerak, Donkers 2007 7 0.104
Wieringa, Verhoef 2007 5 0.044 15 0.006 5 -0.069
Zeelenberg en Pieters 2004 3 0.348
Total 56 0.011 127 0.042 17 0.015 23 -0.053
4.5 Coding
the guidelines set in the discussion. The issues that remained after this coding process were again discussed with an independent advisor, after which the final dataset was agreed upon.
The only variable that had missing values was 'Year of Data Collection'. To create a proxy for these values the average difference between year of publication and data collection was calculated. This average was 4.35 years. The missing data points were filled with the year of publication minus 4 years.
4.6 Moderating Variables
The moderating variables are the variables created from the meta-data of publications on churn that are expected to have an influence on churn driver elasticities. The variables are grouped in five categories: Driver Characteristics, Relationship Characteristics, Market Characteristics, Research Methodology, and Publication Characteristics (see figure 4.3). Table 4.3 gives an overview of all the moderating variables that are included and their expected effect on elasticities.
Table 4.3: Moderating variables of the meta-analysis
Category variable Levels Expected effect on elasticities
Driver Characteristics
Driver Category Critical Incident Perceptions > Behavioral
Relationship Age Demographics unknown
RMIs
Other Behavioral Antecedent Commitment
Involvement
Relationship Quality Satisfaction
Switching Risk
Other Attitude Perception Demographics Other Drivers Relationship Characteristics Relationship Type B2B B2C > B2B B2C Non-contractual > Both >
Contract Type Contractual
Non-Contractual Contractual
Both Market Characteristics
Consultancy + Energy Financial Insurance Manufacturing Retail + +
USA, EUR > Asia, Other Service
Telecommunication Mixed
Region of data collection USA [North America] Europe
Asia Other Research Methodology
Variable Standardized Yes n/a
No
Driver Variable Source Actual n/a
Self-reported
Price Included Yes n/a
No
Year of Data Collection - n/a
Number of Variables - More variables, smaller effect
Dependent Metric Churn Churn > Duration > Retention
Retention
Dependent Type Actual Intention > Actual
Intention
Estimation Method OLS n/a
Logit Probit
Structural Equation Publication Characteristics
Marketing Publication Yes Marketing > Other journals
No
Churn central topic Yes Central topic larger effect
No
Number of Observations - Larger studies smaller effects
Driver Characteristics
The choice of this categorization was made based on the availability and use of these variables in the empirical literature. Demographics have been grouped in one variable. The following levels are included: Critical incident, relationship age, RMIs, other behavioral antecedent, commitment, involvement, relationship quality, satisfaction, switching risk, other attitude perceptions, demographics, other drivers. The literature indicates that churn is most strongly affected by attitude perceptions (Wieringa and Verhoef, 2007; Seiders et al., 2005; Athanassopoulos, 2000). An exception to this evidence may be whether the customer churned or not in the previous period (T-1), as presented by Gustaffson, Johnson and Roos (2005). However, there were no studies with this variable that could be included in this study. Furthermore, Buckinx and Van der Poel (2005) provide evidence that behavioral variables can have a stronger relation to churn than demographics. Therefore we expect that perception drivers will have a stronger relation than behavioral drivers. Because of their situational nature, the demographics and other drivers are expected to have a significant impact on churn.
Relationship Characteristics
Two moderating variables are included that describe the nature of the relationship between the customer and the firm. Relationship type describes whether the customer is a business customer (B2B) or an end-consumer (B2C). Since businesses and end-consumers are expected to value different aspects of a product or service, no clear comparison can be made. However, end-consumers are expected to have a smaller decision making unit (DMU). As businesses are expected to assess their product choices better and with more people, they are expected to be less prone to switching. Indications for this are also provided in Ahtanassopoulos (2000).
The second relationship characteristic is whether a customer is bound by a contract or not. Studies can include contractual, non-contractual or both. Since contractually bound customers have less opportunity to switch, the contractual bound customers are expected churn less than non-contractual customers.
Market Characteristics
bound to have a bigger influence on churn elasticities. These are automobile, consultancy, and retail. There is no firm hypothesis or literature to support this, though.
The region of data collection is expected to have an influence on churn elasticities, as people in Europe or Asia may have a different tendency to switch then for example the North American region. Due to the free trade unions in both North America and Europe, we expect the number of alternatives in these regions to be more abundant, giving customers more choice. In these markets, there has been policy made to make switching service contracts easier as well (Wieringa and Verhoef, 2007). These markets are expected to have a greater influence on churn elasticities.
Research Methodology
To account for the fact that some studies use standardized data and others use unstandardized data, this is added as a moderating variable. The standardization is expected not to change the influence on the dependent variable; although it recodes all the input data to have the same mean, the variance remains.
Analogue to Tellis (1988), Bijmolt et al. (2005) and Kremer (2008), the source of the driver variable is added. It can either be self-reported or actual data received from a database. Since this variable has given mixed results in the past, no expected result can be stated.
Models with a smaller number of variables may lack control variables and are prone to overestimation. The amount of variables in the model, including interaction variables, is added to account for this. A smaller number of variables are expected to have a higher impact on churn elasticities.
Some models include price, or price-level in their model (Bolton, et al., 2006), either to account for changing prices in data that has been collected over longer term, or to be able to assess the influence of discounts. In general we may expect that higher prices lead to higher churn, however this is only relevant in relation to market-level prices. Therefore no expectation can be given whether a price included in the model has an influence on churn elasticities.
example loyalty programs (RMI) may not be as popular now as they were ten years ago and therefore their impact on churn will be less. However, no direction is hypothesized for this effect.
Since not all models use the same dependent metric, we account for this by adding a moderating variable. The variables can be churn or retention. There is no expectation on the influence of this moderating variable.
Churn and retention are not always measured as actual variables; the intention to switch is also measured. There is evidence that people who have the intention to switch do not actually churn (Bansal and Taylor, 2002; Bansal et al., 2005). Therefore it's expected that intention will have a greater influence on churn elasticities than actual churn.
The estimation method is also included as a moderating variable, as done in previous meta-analysis (Tellis, 1988; Bijmolt et al., 2005). This should account for the different estimation methods. We recognize ordinary least squares (OLS), logistic (logit and probit), and structural equation modeling. No prior is specified, though.
Publication Characteristics
Although no unpublished papers are included in this study, publication bias may exist between academic fields of study (Kremer et al., 2008). A dummy moderator variable is added to account for this. Studies published in marketing journals are expected to have a greater bias due to their focus on these topics.
Papers which main topic is about a broader subject than churn, such as Zeelenberg and Pieters (2004), who focus on a broader spectrum on consumer responses to dissatisfaction than solely churn, are expected to have a lower publication bias than studies where the main topic is only churn. A dummy variable is added to indicate whether a publication has churn as the main topic.
Figure 4.3: Moderating variables
4.7 Analysis
A linear model is estimated that uses the moderating effects as independent variables to explain the churn driver elasticities. Following the methodology presented by Bijmolt and Pieters (2001), hierarchical linear models are used for estimation using restricted maximum likelihood estimation (REML). This type of analysis is chosen, since the researched effect sizes come from different studies and therefore have different samples. The intercepts are random estimates per-study, so they can vary randomly. The moderating variables are estimated as direct fixed effects. No nested structure has been defined within these fixed effects.
Figure 4.4 shows a Q-Q plot of the residuals from the statistical analysis. The s-shape of the scattergraph indicates that heteroscedasticity might exist. This is expected to be a result of sampling variability in the observed elasticities. Correcting for heteroscedasticity is not possible, though, since less than half of the studies report standard deviation. However, this is not uncommon in meta-analytic research (Kremer et al., 2008).
Churn Driver Elasticities Driver Characteristics - Driver Category Product Characteristics - Relationship type - Contract type Market Characteristics - Industry type
- Region of data collection
Research Methodology
- Variable Standardized - Driver Source
- Price Included
Figure 4.4: Q-Q plot of residuals
5. Results
This chapter will discuss the results of the meta-analysis. First, descriptive statistics on the overall impact of the dependent variable will be discussed, after which the results of the regression analysis are discussed.
First of all, we look at the dependent variable. The overall mean value of the observed churn elasticities is 0.226, the median 0.001, and standard deviation is 0.307. The frequency distribution of the independent variable is shown in figure 5.1. The distribution is peaked; 51 percent of the elasticities lie between -0.5 and 0.5. The minimum value is -2.048 and the maximum value is 0.926.
Figure 5.1: Frequency distribution of observed churn elasticities
Table 5.2 F-Test of included moderating variables
Variable df F-value p-value* Expected
Driver Category 12 2.441 0.006 Significant
Relationship Type 1 0.134 0.714 Significant
Contract Type 2 11.062 <0.001 Significant
Industry 8 4.266 <0.001 Significant
Region 2 9.360 <0.001 Significant
Variable Standardized 1 1.484 0.225 Insignificant
Driver Variable Source 1 8.471 0.004 Insignificant
Price Included 1 0.000 0.985 Insignificant
Year of Data Collection 1 4.312 0.039 Insignificant
Number of Variables 1 0.517 0.473 Significant
Dependent Metric 1 1.436 0.232 Insignificant
Dependent Type 1 8.604 0.004 Significant
Estimation Method 3 8.864 <0.001 Insignificant
Marketing Publication 1 4.257 0.040 Significant
Churn Central Topic 1 3.542 0.061 Significant
Number of Observations 1 1.140 0.287 Significant
*Bold numbers are significant at the 5% level
The results of each moderator will be discussed individually. Table 5.3 shows the estimation results of the fixed effects of the moderating variables. The table shows the moderating variable and in case of categorical variables, all values that the moderating variable can take. Per variable, the amount of values, the estimate, standard error, and significance are stated3.
5.1 Driver Characteristics
The relationship between the driver category variable and the dependent is as expected. The variable has significant influence on churn driver elasticities (p = 0.006). Moreover, the expected relation between churn and the types of effects are also as hypothesized. Behavioral antecedents have less impact on the dependent variable than perception drivers. Not only are there no significant levels, the absolute impact of behavioral antecedents on the predicted elasticity is also lower than perception variables: the average estimates are 0.027 versus 0.333, respectively.
3
The category level that reports the highest significant churn driver elasticities is affective commitment (β= 0.559, p = <0.001), followed by other attitude perceptions (β= 0.433, p = 0.001), and relationship quality (β = 0.394, p = 0.001). Calculative commitment is significant on the ten percent level (β = 0.283 p = 0.067). The impact of affective commitment is higher than calculative commitment, which is not too surprising (Fullerton, 2003; Verhoef et al., 2002; Bansal et al., 2004). The ambiguous nature of calculative commitment has been recognized in previous research (Jones et al., 2007; Fullerton, 2003). Furthermore, since relationship quality includes variables that are non-churn relationship outcomes of the relationship to the firm (such as word-of-mouth, trust, and quality perception), this significance is not surprising (Wieringa and Verhoef, 2007).
Interestingly, the satisfaction effects show relatively low churn driver elasticities and is insignificant. As recognized in previous studies, the relationship between satisfaction and relationship outcomes can be complex and difficult to detect (Bolton, 1998). Moreover, next to churn, relationship quality is supposed to be an antecedent of satisfaction as well (Verhoef, 2003; Szymanski and Henard, 2001). The inclusion of this level may mask the indirect impact of satisfaction.
The reported demographic effects show relatively high and significant elasticities. This implies that situational and personal influences on churn driver elasticities (Buckinx and Van der Poel, 2005; Mittal and Kamakura, 2001) are relevant when modeling churn.
Table 5.3 Estimation Results
Category variable Levels n Estimate Std. Error t p-value
Driver Characteristics
Driver Category Critical Incident 20 0.104 0.094 1.109 0.269
Relationship Age 7 0.049 0.120 0.407 0.685 RMIs 7 -0.025 0.123 -0.201 0.841 Other Behavioral 22 -0.022 0.090 -0.248 0.804 Affective Commitment 12 0.559 0.144 3.888 0.000 Calculative Commitment 11 0.283 0.154 1.840 0.067 Involvement 1 0.463 0.300 1.543 0.125 Relationship Quality 30 0.394 0.120 3.283 0.001 Satisfaction 18 0.184 0.133 1.388 0.167 Switching Risk 17 0.013 0.099 0.130 0.896
Other Attitude Perception 32 0.433 0.132 3.282 0.001
Other Drivers 23 - - - -
Relationship Characteristics
Relationship Type B2B 26 -0.066 0.179 -0.366 0.714
B2C 197 - - - -
Contract Type Contractual 174 -0.303 0.427 -0.708 0.480
Non-Contractual 28 -2.207 0.778 -2.834 0.005
Both 21 - - - -
Market Characteristics
Industry type Automobile 7 1.475 1.225 1.204 0.230
Consultancy 2 1.561 0.638 2.447 0.015 Energy 25 0.601 0.973 0.618 0.538 Financial 30 -0.991 0.470 -2.111 0.036 Insurance 44 -1.071 0.566 -1.891 0.060 Manufacturing 2 -0.779 0.433 -1.800 0.074 Retail 7 0.994 0.429 2.320 0.021 Service 3 -0.778 0.357 -2.179 0.031 Telecommunications 80 -0.244 0.886 -0.275 0.783 Mixed 23 - - - -
Region of data collection USA [North America] 19 -0.275 0.315 -0.871 0.385
Europe 148 0.292 0.223 1.309 0.192 Asia 54 - - - - Other 2 - - - - Research Methodology Variable Standardized No 193 -0.628 0.515 -1.218 0.225 Yes 30 - - - -
Driver Variable Source Actual 62 0.364 0.125 2.911 0.004
Self-Reported 161 - - - -
Price Included No 195 0.002 0.115 0.019 0.985
Yes 28 - - - -
Year of Data Collection - 0.045 0.022 2.076 0.039
Number of Variables - 0.009 0.012 0.719 0.473
Dependent Metric Churn 139 0.331 0.276 1.198 0.232
Retention 84 - - - -
Dependent type Actual 112 1.320 0.450 2.933 0.004
Intention 111 - - - -
Estimation Method OLS 36 0.611 0.168 3.643 0.000
Logit 124 -0.968 0.360 -2.688 0.008 Probit 10 -0.901 0.577 -1.563 0.120 Structural Equation 53 - - - - Publication Characteristics Marketing Publication No 136 -0.773 0.375 -2.063 0.040 Yes 87 - - - -
Churn central topic No 25 -0.603 0.320 -1.882 0.061
Yes 198 - - - -
5.2 Relationship Characteristics
The majority of the studies analyze a business-to-consumer (B2C) relationship. Of the 223 variables, 197 or 88% are related to a B2C relationship. The other 26 elasticities refer to business-to-business (B2B) relationships. This moderating variable has an insignificant impact on the dependent variable (p = 0.714).
The contract type has a very significant influence on churn driver elasticities (F = 11.062, p < 0.001). The non-contractual relations, 12.56 percent of the effects, have a large impact on the dependent variable (β= -2.207, p = 0.005). This is in line with the hypothesis that non-contractual customers are free to switch and therefore dependent values are expected to be elastic.
5.3 Market Characteristics
The most researched industry type is telecommunications (36%), followed by insurance (20%), and finance (14%). In line with the hypothesis, markets that are less contractually bound industries, such as retail (β= 0.994, p = 0.021) and consultancy (β= 1.561, p = 0.015), have a positive relation to churn driver elasticities, than bound industries such as finance (β= -0.991, p = 0.036). The results show that industry type is a significant effect (F-value = 4.266, p < 0.001).
Interestingly, the region of data collection is significant (F = 9.360, p < 0.001), while none of the levels show a significant relation to the dependent. Most data is collected in Europe (66%) and Asia (24%). The expectation that both North America and Europe have a positive relation to churn cannot be confirmed.
5.4 Research Methodology
matters. No expectation was defined, as previous results were mixed. For this analysis, the source of the variable does have a significant influence (F = 8.604, p = 0.004). The use of actual data instead of self-reported data results in higher churn driver elasticities (β= 0.364, p = 0.004).
For 13% of the effects in the model, a price or price-level variable was added in the model of the analysis. As expected, this variable has an insignificant influence on churn elasticities (p = 0.985). This implies that price or the price-level does not influence churn drivers.
The year of data collection shows a significant relation to churn (F = 4.312, p = 0.039), which implies that churn driver elasticities changes significantly over time. The results show that churn driver elasticities are positively related to the year of data collection.
The number of variables in the analyzed publication is expected to have a negative effect on the dependent variable. The higher the number of variables, the lower the explanatory value of the churn driver variable. The average number of variables in a model is 9. An increase in the number of variables moderator effect specified does not have a significant impact on the dependent variable (p = 0.473).
The switching behavior that is influenced by the churn driver elasticities is for 62 percent measured in terms of churn and for 38 percent measured in terms of retention. To account for possible impact of metric type, the dependent metric moderator effect was added. The moderator variable does not have a significant impact on churn driver elasticities (p = 0.232).
The dependent type effect measures whether the observed churn in the publications has been actual churn, or churn intention. They are equally divided (both 50%) over the publications. The results show that this is a significant effect (F = 8.604, p = 0.004), and that the use of actual data has a positive impact on the driver elasticities. This is not in line with the expectation.
5.5 Publication Characteristics
Of the researched effects, 39 percent were taken from marketing publications. The marketing publication effect is significant (F = 4.257, p = 0.040). The articles that have been published in marketing-related journals show significantly higher effects. This suggests that a marketing publication bias exists and stresses the need for inclusion of a high number of non-marketing related publications in meta-analysis.
The effect that accounts for publications that have churn as a central topic is not significant on the 5 percent level. However with a p-value of 0.061 is significant at the 10 percent level. Of the 223 effects included, 198 have churn as the central topic (89%). This suggests that publication bias may exist, as studies that have churn as the central topic report higher churn driver elasticities.
