R.P.Hars Address: Peizerweg 68-55, 9726 JM, Groningen E-mail: r.p.hars@student.rug.nl Student number: 1871846 Supervisor: Prof. Dr. J.E. Wieringa Date: August 14, 2016

(1)

Discovering the drivers of customer switching behaviour

A meta-analysis

(2)

2

Abstract

Throughout the last decades, a wealth of research has been published on the drivers of customer loyalty, churn and related constructs. One of the main points of criticism on much of this research has been that loyalty is often solely measured as repurchase intentions rather than actual repurchase- and switching behaviour. Calls in the literature have recently been made to generalize what we know about drivers of churn behaviour, given the clear managerial relevance of understanding and predicting such behaviour. Until now however, limited academic attention, limited data availability and methodological limitations have hampered such generalizations.

This study is the first to generalize findings regarding drivers of switching behaviour in a meta-analysis, using 242 effect sizes originating from 36 different studies. I analyse drivers from several categories: perceptions and intentions, product aspects, relationship drivers and demographics. In addition, I analyse a set of study-specific moderators.

The main findings of this study indicate that different drivers have different relationships to switching behaviour. The direction of the obtained relationships between drivers and switching is, in almost all cases, very consistent with previous marketing theory, thus reinforcing earlier findings. The size of the effects, however, is another matter. Specifically, the effect sizes of the perceptual drivers are mostly statistically significant, but surprisingly weak. So weak, in fact, that this directly conflicts with earlier research findings, and with the theoretically central role of several of these constructs –most importantly customer satisfaction– in earlier literature.

Moreover, the results of this study indicate that switching intentions are the one exception to this rule, as they show a relatively large effect size. This finding is encouraging for the validity of earlier work using only switching intentions, and holds promise for the possibility of new, more extensive future studies. Relationship variables are found to also correlate relatively weakly but significantly to switching behaviour, yet their effect size is stronger than for most perceptual variables. Product aspects -specifically price- also show a surprisingly weak correlation to switching behaviour. Finally, not all demographic constructs matter, but age is found to negatively relate to switching, confirming findings in earlier studies.

Finally, the results of this study indicate that methodological differences between studies and situational differences in data collection significantly influence the relationship between switching drivers and outcomes. Most importantly, differences between studies using archival and self-reported data are, in many cases, very large. Journal characteristics however, show no clear effect.

(3)

3

1. Introduction

The question of when and why consumers churn (stop buying from a firm) is a popular topic in the marketing literature. One of the goals of such research is to aid managers in understanding, predicting and -if desired- preventing customers from switching to another product, service or firm. Understanding churn behaviour and improving retention rates is managerially relevant, as an increase in customer retention rates can have a direct impact on a firm’s profits (Bolton et al. 2004).

Customer churn behaviour is an important aspect of several theoretical models in the marketing literature, such as the CUSAMS framework (Bolton et al. 2004). In addition, many papers, especially in the marketing literature, have attempted to relate churn-related constructs to sets of potential drivers (e.g. Wirtz et al. 2014). In general however, the study of churn is of a surprisingly fragmented nature. It can be studied within many theoretical frameworks, with many different types of statistical models, and has been studied using many different conceptualizations.

Research objective and contribution to literature

Given the subject’s managerial and theoretical relevance yet fragmented nature, it is unsurprising that there have been calls in the literature to generalize our knowledge concerning the drivers of churn (Verhoef et al. 2007). What is surprising, however, is that these calls have still not been fully answered. Although recent meta-analyses have attempted to generalize drivers of related concepts such as loyalty (Palmatier et al. 2006, Pick & Eisend 2014, Watson et al. 2015), these studies have serious limitations, which I will discuss below. Thus, in contrast to these earlier studies, my research goal is:

To generalize knowledge concerning the drivers of customer churn behaviour

Aside from addressing the call to generalization, this study makes several contributions to the literature. First, as previously mentioned, earlier meta-analytic studies have considered concepts related to churn, such as loyalty (Palmatier et al. 2006, Watson et al. 2015). However, the usual operationalization of the loyalty construct involves measuring attitudes and intentions rather than actual behaviour. While both should certainly be related, this assumption has not been thoroughly considered in previous meta-analyses. At the same time, the strength of the relationship between switching intentions and behaviour is a much debated subject (Chandon et al. 2005). Therefore, the managerial relevance of these previous meta-analyses and much of the underlying literature directly depend on how strong this relationship is. In this analysis, I will address this question.

(6)

6

Third, a meta-analysis can directly contribute to debates concerning the managerial and theoretical relevance of individual constructs. For example, in the marketing literature, customer satisfaction is often considered to be one of the main drivers of customer retention, loyalty, market share, and related measures. However, there is serious disagreement about how central, relevant and important customer satisfaction really is to subsequent retention, repeat purchases, and so forth (Anderson et al. 2004, Bolton 1998, Rego et al. 2013). Performing a meta-analysis helps to answer such questions.

Finally, as I will argue below, there is a good reason why earlier meta-analyses have not thoroughly considered churn behaviour. Due to the fragmented and methodologically diverse nature of the subject, churn is an inherently difficult topic for meta-analyses to address. To make matters worse, the currently available meta-analytic methodology is not able to synthesize results from a diverse array of statistical models that are commonly used to study churn, such as logistic and probit regressions. Therefore, studying churn requires a methodological contribution which does allow for such study.

Outline

(7)

7

2. Theoretical model

Theories of churn/loyalty and conceptual definitions

Throughout this thesis, I define churn as follows. A customer churns when they decide to stop purchasing a product or service from a certain firm, brand, dealer or provider. In contrast, retention is a customer’s decision to keep purchasing a product or service from a certain firm, brand, dealer or provider. Therefore, I define churn and retention only as the decision to stop or keep purchasing in itself. It may therefore be different from constructs such as repeat purchase, in which the absolute number of purchases itself is often considered (e.g. Sharp & Sharp 1997). It is also different from (but related to) the construct of loyalty. Earlier literature has defined loyalty as “a deeply held commitment to rebuy or re-patronize a preferred product/service consistently in the future, thereby causing repetitive same-brand or same brand-set purchasing, despite situational influences and marketing efforts having the potential to cause switching behaviour” (Oliver 1999, p34).

This relationship between churn and loyalty is important in our theoretical discussion. Although there are theoretical frameworks that directly focus on churn, most frameworks have a broader character, and focus on loyalty or similar, related constructs. As drivers of customer loyalty may also be drivers of churn, such models are still informative in the current setting. In the section below, I will discuss four of the most important and popular frameworks in the literature, which I believe provide a good overview of different potential drivers of churn. In addition, there are many related theories which I do not discuss here. A good survey of the literature can be found in Toufaily et al. (2013).

Dick & Basu’s loyalty framework

(8)

8

In general, Dick and Basu identify three types of loyalty antecedents. First, they distinguish so-called cognitive antecedents. Examples of such antecedents are accessibility (how easily a brand attitude can be retrieved from memory), confidence (certainty of one’s attitude), and clarity (how well-defined one’s attitude is). Affective antecedents, in contrast, relate more to emotions (e.g. liking a store’s personnel) and to satisfaction (defined as an emotion arising from the matching of perceived performance and expectations). Third, the authors distinguish conative antecedents Examples of these are switching costs (one-time costs facing the buyer when switching), sunk costs (costs that have already been incurred, these have been found to increase repeat patronage), and future expectations.

In general, the main contributions of their model lies in that it distinguishes loyalty attitudes from behaviour, in that it introduces situational moderators (e.g. opportunity, discounts), and that it identifies at least several antecedents of these attitudes (e.g. satisfaction and switching costs), which have also been considered in subsequent studies.

Figure 1. Dick & Basu (1994) framework

Churn in relationship marketing: commitment and trust

(9)

9

reasoning (e.g. commitment due to switching costs). In contrast, affective commitment is considered to be a more emotional form of commitment. Thus, both forms of commitment are closely related to the affective antecedents and situational constraints in Dick & Basu’s framework.

Figure 2. Morgan & Hunt’s (1994) framework

The CUSAMS framework

A framework that is closely associated with the relationship marketing literature is the so-called ‘customer asset management of services’ (CUSAMS) framework (Bolton et al. 2004). The framework (figure 3) is a conceptual model that shows how relationship marketing instruments can influence three different types of consumer behaviour through influencing customers’ perceptions of their relationship with a firm. In terms of behaviour, the framework distinguishes relationship length (e.g. how many years one is a customer, very closely related to churn), relationship depth (frequency of product/service usage), and relationship breadth (the extent of cross-buying of different types of products from the same provider).1

In contrast to the previous frameworks, this model is more specific in discriminating between the effects of different marketing instruments. Six instruments are distinguished, namely the price of a product or service, promotions (e.g. mailings), relationship marketing programs (e.g. loyalty programs), different distribution channels, effects of advertising (e.g. aimed at creating brand awareness) and programs aimed at increasing service quality. Such instruments can influence both perceptions and behaviour. Three main perceptions are included in the model: price perceptions, satisfaction and commitment. As can be seen, these perceptual variables are very similar to the variables that are encountered in the earlier frameworks.

1_{The framework also discusses how length, depth and breadth relate to revenues and lifetime value. I will}

(10)

10

Most importantly, the model holds that different instruments may influence customer perceptions and behaviour in different ways. For example, whereas price promotions may be effective on the short term, a successful service quality program can permanently increase service quality. Improved service quality, in turn, may have a permanently positive effect on many forms of customer behaviour, including retention (Bolton et al. 2004).

Moreover, the model proposes several potential moderators. In general, these moderators are related to product/service aspects and to competitive aspects. Examples are switching costs, competitive intensity, the extent of consumer product involvement, and perceived risk (uncertainty about product/service quality). In other words, besides discussing marketing instruments as potential drivers of churn, it introduces two clear categories of moderators.

The framework also distinguishes three aspects of a customer-firm relationship, namely relationship length, breadth and depth. However, these are not discussed as antecedents of retention, but as outcome decisions. Other studies, in contrast, have treated these three variables as antecedents of retention (Verhoef 2003, Dawes 2009). Doing so can be valuable. For example, a common issue for firms in studying churn behaviour is that perceptual variables are not easily available, and that these can be very costly to obtain. Therefore, if it is indeed the case that relationship marketing instruments can lead to increased relationship length, breadth and depth by influencing perceptions, then including these behavioural variables may be useful proxies for positive perceptions and/or other drivers of loyalty (Verhoef 2003 argues a similar case). Therefore, I will also use this approach, and compare perceptual drivers to these customer-firm relationship aspects in terms of effect size and directionality in the results section.

(11)

11

Keaveney’s grounded model of switching

A final theoretical framework departs from previous models in three important ways. First, whereas the previous frameworks mainly describe positive outcomes (loyalty, customer lifetime), Keaveney’s (1995) model is directly concerned with a negative outcome: switching. In addition, it focuses directly on behavioural outcomes, not on broader dependent variables such as loyalty. Moreover, this framework has been established using qualitative data. In essence, Keaveney et al. simply asked consumers why they switch firms, and subsequently generalized their answers into exhaustive and mutually exclusive categories. As a result, many variables in Keaveney’s framework are of a more incremental, behavioural and short-term nature.

Their framework contains eight categories with reasons for consumers to switch firms. These are: issues related to pricing (e.g. deceptive or unfair pricing), inconvenience (e.g. due to poor service quality resulting in a long waiting time), core service failures (e.g. internet outage), failures during service encounters (e.g. rude staff), a company’s poor response to a service failure (e.g. complaints not addressed), competitive aspects, ethical issues (e.g. conflicts of interest), and involuntary switching (e.g. having moved). Although Keaveney et al. do not discuss how these factors relate to perceptions, they factors are still useful in understanding and predicting churn behaviour. Moreover, other academics have actually used this framework in empirical work (e.g. Athanassopoulos 2000). Therefore, these drivers are very useful to take into account in a general empirical model describing customer churn.

Generalizing across theories: a conceptual framework of churn behaviour

Based on the conceptual models that I have previously discussed, I will make a distinction between four very general categories of drivers. These categories are broadly similar to those discussed in Buckinx & Van den Poel (2005). While it would be very valuable to also discuss mediating and moderating relationships, this cannot be done due to limitations. For this reason, the framework remains relatively simple, theoretically speaking.

Many of the previous frameworks contain perceptions, attitudes and intentions. Most importantly, previous frameworks have distinguished satisfaction, convenience, switching costs, perceived quality, loyalty intentions, commitment and involvement. I summarize these into the general category of perceptual variables, as they all relate closely to how an individual thinks, feels, perceives, or aims to act.

(12)

12

Third, these frameworks distinguish variables that are related to the relationship between a customer and a firm. I will call this category ‘relationship aspects’. Here, I include variables such as relationship length, breadth or width, relationship marketing efforts, and critical incidents. To be sure, several variables in the relationship marketing category could also be called perceptual and/or attitudinal variables (e.g. commitment, trust). To clearly distinguish between both categories, I only include variables in the relationship category if they are not perceptions or attitudes.

Finally, while not present in the previous frameworks, many primary studies also include demographic covariates (e.g. gender). Although there are usually no theoretical reasons to expect differences between different demographical variables, there is an important exception: age. Lambert-Pandraud et al. (2005) argue that older consumers tend to switch products and firms less due to cognitive decline and change aversion.

In general, there has not been enough empirical work to include all the previously mentioned variables in the analysis. Based on the fact that they have been sufficiently empirically studied, I include the following variables in these categories. In the demographics category, I include gender and age. I include price and usage intensity as product-related variables. The relationship category consists of relationship breadth, relationship length and the effect of loyalty programs. Finally, perceived convenience, perceived quality, customer satisfaction, perceived switching costs, perceived risk of switching and loyalty intentions are included in the model as perceptual variables. The definitions of these constructs are available in table 1.

To be sure, this means that many potentially important constructs such as critical incidents, trust, commitment, other marketing instruments, promotions, distribution channels, competitive intensity and involvement are all not included in this study. This is due to a lack of data. While most of these constructs have been the subject of earlier studies, there are not enough studies available to summarize these effects in a consistent and satisfactory method. As a minimum criterion for inclusion, I have set the number of samples that had to be available at three. The constructs above either have not been commonly included in churn models, or they have not been included in a fashion consistent enough to make the effects comparable.2_{The full conceptual model for this study can be found in figure 4. I will now discuss the}

relationships I expect between the individual constructs and switching behaviour. I give an overview of these hypotheses in table 2.

2_{For example, income is commonly included in churn models as a categorical variable. However, the exact}

(13)

(14)

14

Table 1. Conceptual definitions

Variable Definition/measurement

Age Age in number of years, only measured as a continuous variable

Convenience The ability to do business with a firm without difficulty. Usually asked in relation to: physical location, hours of operation or waiting time (Keaveney 1995)

Gender Dummy variable indicating whether a customer is male (=1) or female (=0)

Loyalty Programs Whether a customer is a member of a loyalty program, which is "An integrated system of marketing actions that aims to make member customers more loyal" (Leenheer et al. 2007, p32)

Perceived price A customer's perceived fairness/positivity of the price paid for a firm's products or services (Bolton & Lemon 1999, p173). Also referred to as payment equity

Price

The amount of money a customer has to pay for a product or service, either in absolute terms, or per use. Measured in relation to other customers of the same firm, in relation to competitors, or as a change in comparison to earlier levels

Quality

The perceived quality of a product or service, based on reliability, responsiveness, assurance, tangibles and empathy of the service provider (Parasuraman et al. 1988). Usually, but not always measured as performance in comparison to expectations (Cf. Cronin & Taylor 1992).

Relationship breadth The extent to which a customer buys additional (different) products or services from the same company over time (Blattberg et al. 2001)

Relationship length The duration of an (uninterrupted) relationship between a customer and a firm, based on their purchase behaviour. Usually measured in months or years

Switching intentions

A customer's stated probability that they will switch from their current firm to an alternative firm, or stop purchasing a product or service (Bansal & Taylor 1999). Opposite of repurchase/loyalty intentions, which refer to the a customer's stated probability of repurchasing, recommending or increasing their share with a firm's products or services (Bolton et al. 2000)

Satisfaction

"A customer's response to the evaluation of the perceived discrepancy between prior expectations (or some other performance norm) and the actual performance of the product as perceived after its consumption" (Tse & Wilton 1988, p204). Can also include expectations based on future use (Lemon et al. 2002)

Switching costs

"The onetime costs that customers associate with the process of switching from one provider to another", such as search costs, transaction costs, loyalty discounts, emotional cost, cognitive effort (Burnham et al. 2003,p110)

Switching risk

The extent of or aversion to risk associated with switching to a new vendor (Capraro et al. 2003). Has also been considered as a sub-dimension of switching costs (Burnham et al. 2003). I separately measure it due to its common occurrence and potentially more specific nature.

Usage intensity and frequency

(15)

15

Hypotheses

Age

I expect a negative relationship between a customer’s age and their switching probability. On the basis of relatively well-accepted gerontological findings, Lambert-Pandraud et al. (2005) argue that aging is related to biological processes that cause cognitive decline and change aversion. Cognitive decline may make comparing alternative products and service more difficult, and therefore essentially increases switching costs (see table 1 for the definition). Moreover, change aversion is related to a decrease in switching by definition. For both reasons, I expect a negative relationship between a customer’s age and their switching propensity.

Relationship length

There are several reasons to expect a negative association between relationship length and switching behaviour. First, Verhoef (2003) makes the point that past behaviour will in general be indicative of future behaviour. Therefore, if a customer (given enough time) has not switched companies, it would be reasonable to not expect them to switch in the foreseeable future. In addition, if customers remain at a single firm for a longer period of time, they have the additional advantage of having lower transaction costs. This is the case, because a renegotiation of contracts and terms does not have to take place (Rese 2003). Furthermore, maintaining a relationship saves the customer effort that would otherwise be spent searching for alternatives, and also means the customer does not have to incur switching costs.

Relationship breadth

I expect a negative association between relationship breadth and switching, such that the more additional products a customer purchases, the lower the odds of switching firms are. Most importantly, many firms offer economic benefits (e.g. price reductions) to customers who buy multiple products rather than a single product (Bolton et al. 2004). This added utility of being a product essentially gives a customer more ‘value for money’. Therefore, it can make a product more attractive in comparison to competing products, in turn making switching less attractive. In addition, given a customer’s knowledge of and positive experience with a firm, buying more products from the same firm may be associated with a lower risk tolerance. Given that customers already know the firm that provides the product or service, buying alternative products or services from the same firm is associated with less risk. As switching is inherently risky given limited knowledge of competing options (e.g. Rese et al. 2003), I expect such risk aversion to be negatively related to switching behaviour. For both these reasons, I expect relationship breadth to be negatively related to switching.

Loyalty program membership

(16)

16

recommendation behaviour. Such programs are usually associated with economic or other benefits for customers (e.g. discounts based on repeat purchase) which would otherwise be unavailable. This means that the utility of being a customer of a firm increases (Bijmolt & Leenheer 2002) in comparison to competitors once a customer takes part in a loyalty program. Given these additional benefits, alternative options may relatively less attractive. Therefore, have less reason to switch to competitors.

Therefore, I expect a negative relationship between loyalty program membership and switching. However, previous literature has offered some conflicting findings. For example, Bolton et al. (2000) have found that loyalty program membership generally increased customer spending, but did not increase retention. Others (e.g. Verhoef et al. 2001) report positive effects on retention. Melnyk & Bijmolt (2015), have suggested that characteristics of the competitive environment and of the loyalty program itself are important. In general however, I still expect a negative relationship between membership and churn, but also expect a significant amount of heterogeneity between studies, as such program- or context-specific moderators may be important.

Usage intensity/frequency

I expect a negative relationship between usage frequency and switching behaviour. There are several arguments for these expectations. First, literature suggests that a higher product usage frequency is associated with stronger attitudes about a product (Keaveney et al. 2001). Given that a customer still uses a product at the time of measurement, Keaveny argues that we can infer that these attitudes will in most cases be positive. After all, customers with strongly negative attitudes can be expected to have already churned in the past (Keaveney et al. 2001). Previous empirical work (e.g. Bolton et al. 1999), has found that increased product usage is associated with increased satisfaction, which is commonly associated with a decrease in switching.

In addition, the continued and increasingly intensive use of a product may also increase switching costs. As argued by Keaveney et al. (2001), individuals consider learning how to use a product as a relatively effortful process. The more intensive and often one uses a product, the more familiar one becomes with the product. The more familiar one becomes with a product, the easier it is to gain the notion that one has developed skills in using this specific product which may be non-transferable to other products. As becoming familiar with competitive products requires additional effort, increased product use can increase switching costs, which in turn can decrease churn, as argued below.

Satisfaction

(17)

17

should have to switch offerings. Satisfaction has also been positively associated with (self-reported) loyalty in earlier meta-analyses, such as Szymanski & Henard (2001) and Watson et al. (2015).

Switching costs and risk

Switching costs are one-time costs that are incurred on switching. For example, when switching services, customers may be required to sign contracts or fill in forms. Before finding a new service provider, a customer may need to expend effort in comparing products, reading reviews and so forth. In addition, switching may also be associated with financial costs (Hu et al. 2006). Aside from such search- and transaction costs, a customer also loses any benefits they may have had at their old firm, for example additional benefits due to loyalty programs. The higher such costs, the more expensive and effortful switching becomes. Therefore, I expect a negative relationship between switching costs and switching behaviour. In addition, because switching risk can be considered a specific dimension of switching costs, I also expect a negative relationship between switching risk and churn.

Convenience

As can be seen in table 1, I define convenience as referring to the ease with which a product or service can be used. The construct is related to aspects such as a service provider’s location, hours of operation, and speed of service. A lower convenience means that a product or service becomes more difficult to use, which should negatively influence the utility a consumer has for it. After all, fewer hours of operation or an inconvenient physical location means that a customer now has to expend more time, effort and money to obtain the same service. This, in turn, should increase the relative attractiveness of competitive offers, decrease relative switching costs, and therefore lead to more switching behaviour.

Switching intentions

I expect a positive relationship between switching intentions and switching behaviour. Although there has been a significant amount of debate in the literature about the extent to which customers’ intentions actually correspond to their behaviour (Anderson et al. 2004, Bolton 1998, 2000, Rego et al. 2013), there is –to the best of my knowledge– no debate regarding the general notion that intentions should be positively related to behaviour. In addition, it is common in psychological theory to directly relate intentions to behaviour (Ajzen 1991).

Perceived product/service quality

(18)

18

In contrast, low service quality is directly related to service failures. Such service failures are well-known triggers for service switching, as they may dramatically decrease the usability and thus relative utility of a service (Keaveney 1995). In other words, high service quality should by definition be related to a lower occurrence (or at least lower impact of) service failures, which gives consumers less reason to switch.

Price and perceived price

Price perceptions are an important part of the perceived utility of a product or service. The more positive perceived prices are, the higher the relative benefits of the products are in comparison to alternative options. This, in turn, should give customers less reason for switching products or services. In addition, empirical studies (e.g. Bolton et al. 2001) have demonstrated that perceived (price) fairness is associated with a longer relationship length, and a higher relationship depth, which I previously argued to be negatively related to churn.

Therefore, I expect that more positive price perceptions are negatively related to switching behaviour. To the extent that price perceptions are directly influenced by the product’s actual price, I expect that for most products, a product’s actual price in comparison to the price of relevant alternatives is positively related to switching behaviour. My reasoning for this is the same as for perceived prices. If two identical products differ only in price, I expect customers to generally have a higher utility for the lower priced product, which increases the product’s relative benefits in comparison to alternative options.3

Moderators

Aside from the available constructs, I will include moderator variables in the analysis. To be sure, these are not theoretical constructs. Rather, it is common practice in meta-analysis to include study-specific variables (e.g. the type of firm studied, the type of model used, and so forth) that may explain heterogeneity between findings from different studies. I will discuss these variables in the methodology section rather than here for two reasons. First, due to their a-theoretical nature, they are a more methodological than theoretical aspect. Moreover, I will give an overview of related meta-analyses in the next section. These studies contain examples of which moderator variables are commonly included in such meta-analyses. Therefore, they have directly influenced my choice for specific moderators, and therefore should also be considered before discussing my own choice for moderating variables.

3_{I do not go into detail here about possible exceptions, such as luxury products (not in the sample of studies), or}

(19)

19

Table 2. Hypotheses

Variable Expected relationship with churn

(20)

20

Calls for generalization, and earlier meta-analyses: previous findings and limitations

In their 2007 paper, Verhoef et al. argue that with the recent increase in data available to firms, using such data to predict and increase the lifetime value of a firm’s customers (i.e. the profitability associated with the entire future relationship with a customer) has become much more practically plausible. Therefore, research that allows firms to make such predictions more accurate and relevant has gained increased theoretical and practical relevance, and has become much more common. Increasing our knowledge regarding the drivers of customer retention directly ties in with this research, as predicting customer retention is an essential part of predicting customer lifetime value.

The authors give a very brief but thorough qualitative overview of some important debates related to customer retention. They make three observations that can be compared to the results in this meta-analysis. Most importantly, Verhoef et al. (2007) note that customer satisfaction is the most frequently studied antecedent of churn. However, the relationship between customer satisfaction and retention has been thoroughly debated, and the authors note the evidence is either weak or inconclusive. Moreover, the authors note that the evidence for an effect of price perceptions on customer retention is also relatively weak. Third, they find that most previous studies have found a significant effect for loyalty programs, but that such effect may be inflated due to self-selection effects.

In their review, the authors (p62) define four questions that are important to answer to further progress our knowledge of customer retention. These are (reproduced here almost verbatim):

1: Can a general model explaining customer retention be developed? Which models (and thus drivers) best explain retention behaviour?

2: How does the effect of determinants of customer retention differ between industries and market environments?

3: How does the effect of determinants of customer retention differ between countries? Is there a role for economic and cultural characteristics?

(21)

21

There are multiple meta-analyses that can form an important basis for such generalizations. These studies are closely related to the current study. I will briefly discuss four recent meta-analyses for several reasons. First, these show which of the questions in Verhoef et al. (2007) have and have not been answered. Therefore, they place the current work in a broader context. Second, although not always directly related to retention, these studies still give an indication of the type of results we can expect. Therefore, discussing these analyses also allows a better interpretation of the findings in this analysis as being consistent or divergent with earlier work. Finally, most meta-analyses investigate a set of moderating variables which can influence the strength of the results obtained. Studying earlier meta-analyses gives a useful indication for what moderators could be included. I discuss several aspects per study: the constructs included, the size of the effect sizes, and the included moderators. Finally, I will briefly discuss how these studies relate to the current study. Most of these studies use the correlation coefficient to discuss the size of a relationship between two variables.

De Matos & Rossi 2008: Word of mouth

A tangentially related but nonetheless interesting meta-analysis has been conducted by de Matos & Rossi (2008). Their conceptual model can be found in figure 5. The authors were interested in investigating word of mouth rather than churn as a dependent variable. They define this as “informal communications directed at other consumers about the ownership, usage or characteristics of particular goods and services and/or their sellers” (de Matos & Rossi 2008, p578).

(22)

22

Figure 5. De Matos & Rossi (2008): Word of mouth

Palmatier et al. 2006: Relationship marketing effectiveness

One of the broadest meta-analyses on the effectiveness of relationship marketing factors to date has been conducted by Palmatier et al. (2006). The study’s conceptual model is displayed in figure 6. The study distinguishes several antecedents of relationship marketing outcomes, several mediators (similar to variables discussed in the theoretical model above) and several outcomes. The authors report on 94 different studies. I focus on loyalty and so-called seller objective performance (sales, profit, tobin’s Q, but not retention) here, as these dependent variables are of main interest.

The reported relationship between the mediating variables in their study (which are closest to the variables in our models, e.g. relationship satisfaction, trust, commitment) and loyalty is very similar across constructs. The 95% confidence interval for most effect sizes is between 0.4-0.6. Interestingly, when considering behavioural measures (seller objective performance, e.g. sales levels), the correlations between the mediators and the dependent variables are lower, usually between 0.2-0.4. This finding is therefore similar to the results obtained in the de Matos & Rossi (2008) study.

(23)

23

The study’s relevance to this thesis is limited, for several reasons. First, the number of variables that are directly related to loyalty (which seems to come closest to churn) in their model is limited to four. Therefore, the model is relatively small. Second, in their loyalty definition, the authors make no distinction between attitudes, intentions, self-reported behaviour and more objective forms of behaviour, judging by their definitions and the list of representative papers that were provided in the study.4_{This is}

a theoretically important omission. As has become clear from the earlier discussion and the previous two meta-analyses, the correlation between attitudes, intentions and any form of actual behaviour (not necessarily churn) is generally lower than the correlation between perceptions and attitudes on the one side, and intentions on the other. In other words, their findings cannot simply be generalized to switching behaviour.

Figure 6. Conceptual model, Palmatier et al. 2006

(24)

24

Watson et al. 2015: relationship marketing

The fact that Palmatier et al. (2006) do not distinguish between attitudes and behaviour, as is theoretically important, has also been noted in a very recent meta-analysis by Watson et al. (2015). The authors note that one of their main contributions in comparison to the Palmatier study is to actually make the distinction between attitudinal and behavioural loyalty. Specifically, the authors distinguish (figure 7) four independent variables: commitment, trust, satisfaction and loyalty incentives. In addition, they distinguish two dependent variables, namely word of mouth and (sales) performance. Moreover, they investigate the extent to which attitudinal and behavioural loyalty mediate the relationships between these variables.

In terms of definitions, the authors note (p793) that attitudinal loyalty is “a cognition or pleasurable fulfilment favouring one entity such as a firm, its brand, its salespersons or its offerings”. In contrast, behavioural loyalty is defined as “repeated purchases that stem from a conation or action orientation involving a ‘readiness to act’ favouring one entity”. While attitudinal loyalty equates to affect, preference and warmth in their study, behavioural loyalty equates to purchases, repurchases, repurchase intentions, retention and return behaviour.

For our purposes, this seems to be a clear an improvement on the Palmatier et al. (2006) study. However, their behavioural loyalty construct is still very broad, as it includes both behavioural intentions and different forms of behaviour. Moreover, the primary studies that were included seem to address intentions much more than behaviour. In fact, I have investigated the list of studies included by the authors, and there seems to be no similarity at all between the behavioural studies cited in their paper, and the studies included in this thesis.5

In terms of results, the independent variables correlate to attitudinal loyalty with a very comparable effect size to the Palmatier et al. (2006) and the de Matos and Rossi (2008) papers. The confidence intervals of most effect sizes are between 0.4-0.7, except for loyalty incentives (-0.06-0.58). These correlations between the independent variables and loyalty were generally similar for the attitudinal and behavioural loyalty items. Interestingly, the correlation between attitudinal loyalty and objective performance –which is closer to actual behaviour– was low (0.11-0.38) whereas behavioural loyalty correlated more strongly with objective performance (r=0.26-0.51).

The moderators included by the authors seem to also confirm the crucial distinction that must be made between attitudes, intentions and actual behaviour. For example, the more attitudinal measures that were included in a construct, the weaker the relationship with sales performance became. In addition, including more forward looking items (i.e. future intentions, plans), rather than backward-looking items

5_{That is, zero of the studies that I include here were used by the authors. However, their list of literature does}

(25)

25

(which may be more objective: e.g. past behaviour) weakened the relationship with objective performance measures. To be sure, this is a correlation with sales-related variables rather than churn, but the two constructs may well be related.

Finally, the authors found significant differences in the size of the correlation coefficients between business and consumer markets, and in year of publication. Although several other moderators were also taken into account (e.g. brand versus non-brand loyalty, whether studies were unpublished versus published and journal quality), no effects for these were detected.

Figure 7. Conceptual model, Watson et al. 2015

Pick & Eisend 2014: Switching

The meta-analysis that comes closest to this thesis in terms of the dependent variable (see figure 8 below) is a study by Pick & Eisend (2014). In the analyses, the authors are specifically interested in switching, and they study both intentions and behaviour. Although their conceptual model is different from the model in this thesis, there are some similarities. For example, both include a quality measure.

The reported effect sizes are very different from the effect sizes in the previous meta-analyses. While many of the independent variables in Pick & Eisend’s conceptual model are different, this does raise the question of whether these different effect sizes are related to a different set of independent variables, or mostly to a different dependent variable.

The correlation between the study’s independent variables and switching is lower than the correlations between the independent variables and loyalty and/or word of mouth in other meta-analyses. At the very least, their estimates have a wider confidence interval. For example, the correlation between perceived quality and switching intentions is between -0.408 and -0.201 (95% confidence interval). The effect size of switching costs is between -0.162 and -0.016.

(26)

26

that the correlation between most variables and behaviour are much weaker than between most variables and intentions, which corresponds to results in other meta-analyses. For example, the correlation between critical incidents and switching intentions is 0.447, yet only 0.028 for switching behaviour. Moreover, when considering the list of studies the authors have used, one things becomes immediately obvious: The paper suffers from an extremely low number of primary studies that measure behaviour rather than intention. For example, for the correlations between quality and switching behaviour, seller investments and switching behaviour, incidents and switching behaviour, involvement, competition and switching behaviour, the number of effect sizes available are either 2, or even 1. While their two remaining variables (perceived quality, switching costs) fare better in terms of the number of effects (12 for switching costs, 15 for quality), the number of participants still suggest that these effects originated from a very low number of studies. I cannot exactly determine how few. However, by investigating their list of included studies, I estimate that the number of studies these effects are based on are probably 3 or fewer.

In other words, while initially appearing otherwise, this study’s findings cannot be robust, given the very limited number of studies. To be sure, this is due to the fact that many switching behaviour studies do not report bivariate correlation coefficients or similar statistics, which the authors use this an effect size measure in their study. Therefore, the issue here is of a more methodological nature: most studies involving churn report models (e.g. logistic regression) for which it is unclear how to synthesize results. I will address this in the next chapter.

(27)

27

Conclusions from previous meta-analyses

Having discussed these previous analyses, we can draw several conclusions. First, most meta-analyses have considered loyalty and related constructs, but have paid limited attention to actual behaviour, especially in a churn context. In addition, it is surprising to see that the effect size measures for most broad, loyalty-related (and often survey-based) statistics are relatively similar across most analyses, even when different definitions and dependent variables are studied.

(28)

28

3. Methodology

In this chapter, I will discuss the methodology that I have used. Where more technical, lengthy and/or relatively unimportant details are concerned, I will refer the reader to an appropriate appendix, or other works on the subject. The chapter is divided into multiple sections. I will first discus the search procedure that I have followed to obtain an initial set of studies. Next, I will list the study inclusion criteria that have been used to limit the number of observations. Third, I discuss coding decisions. In the fourth section, I discuss which moderators are included in this study. The final sections concern the effect size measure that I will use, the econometric model used to analyse the data, and the exact modelling procedure that I will follow.

Search procedure

As this thesis is part of a larger ongoing project, I will discuss the search procedure for the project as a whole here. First, we have established a set of relevant keywords such as ‘churn’, ‘retention’, ‘defection’, ‘turnover’ and several other keywords and variations thereof. The full list of keywords can be found in appendix 1. In addition, a list of potentially relevant journals has been established, and has been used to initially limit the number of publications found in the database to a set of more relevant papers of potentially higher quality. In terms of databases, we have searched for our keywords on ISI Web of Science, Sciencedirect and Ebscohost. We followed this procedure for multiple databases because not every journal was contained in every database, and the search algorithms per database were potentially different.

(29)

29

In addition, in a later stage of the search, I encountered several of the previously-discussed meta-analyses. To the extent that the these analyses reported on the primary studies that were used as data, these have also been searched. For this thesis, I only use the publications from our initial search, from the related meta-analyses that were previously discussed, and from a sub-sample of the papers in the grey literature.6_{For a full overview of the included papers, see appendix 3.}

Inclusion criteria

The inclusion criteria in this meta-analysis are broad in certain senses, but relatively restrictive in others. Studies must meet several inclusion criteria to be considered for the analysis. First, I only consider retention/churn/switching as a yes/no decision, and only in a lost-for-good situation. This criterion excludes many types of studies from the analysis, even when they are clearly addressing churn. For example, I exclude studies that report on churn with hazard models (e.g. Armelini et al. 2015, Jamal & Bucklin 2006), studies that report on churn as a continuous variable (e.g. number of months a customer was retained out of a total number of months, Gustaffson et al. 2006), studies that report on churn as ordered dependent variable (e.g. Echambadi et al. 2013) and migration models that infer retention/churn from the number of purchases that customers have made (e.g. Knox & Van Oest 2014). The reason for this choice is mainly methodological. To the best of my knowledge, until recently, there have been no effect size measures for any other situation than simple bivariate and linear relationships. Although advances have recently been made in the case of meta-analysing coefficients from more extensive linear models, this is not the case for other models (Aloe 2015, Aloe & Becker 2012, Peterson & Brown 2005, Wu & Becker 2012). In this thesis, I combine results from logistic models, probit models and simpler bivariate correlations. However, I am unaware of a method to combine estimates from these models with estimates from other types of models, such as the aforementioned. Therefore, they are excluded from the analysis.

In addition, I have excluded studies which measure intentions (e.g. intention to remain a customer) rather than behaviour as a dependent variable. Most importantly, this makes the set of included studies relatively similar, both in terms of dependent variables and in terms of methodology. I believe this aids the strength of the conclusions. Second, the number of ‘intention’ studies is very large.7_{Therefore, this}

is outside of the scope of this thesis.

Studies that were not easily accessible were excluded from the sample. After having included a study in the search database, if it could not be downloaded/viewed through the university’s access to the

6_{The coding of the papers in the grey literature is still an ongoing project. Thus I have included the papers that}

were available to me at this stage in the project.

7_{I do not know the exact number. However, I have read all studies that we found in our initial (non-grey)}

(30)

30

aforementioned databases and the study could not easily be obtained e.g. through search engines and/or the university library, it was excluded from the analysis.

Furthermore, the study had to have at least one author affiliated to an academic institution. This criterion was included as a minimum quality measure, and to limit the number of studies considered. While academics are not necessarily more objective than others, their reputation partially depends on their publications, which may result in a higher accuracy. In addition, this has the advantage of excluding e.g. bachelor and master-theses, somewhat limiting our sample.

An additional and rather obvious criterion is that only quantitative studies are included. From these studies, it had to be clear which variables were actually being measured, based on the provided definitions or references to other papers. This is necessary to avoid a common pitfall for meta-analysts where similarly-named constructs are actually dissimilar in their definition or measurement, whereas differently-named constructs may actually be similar in their definition or measurement (Landis 2013). In case of doubt, papers were excluded from the analysis.

Next, I only consider profit seeking firms, as behaviour towards non-profits may be different from for-profit firms, as they may include ideological motivations. In a similar vein, I restrict the sample to consumer markets only, as these are distinct from business-to-business markets, which may involve very different mechanisms (Watson et al. 2015).

Moreover, the study needs to contain variables which relate to theoretical constructs, as many churn studies contain very context-specific variables which are not easy to generalize. For example, it is common to encounter dummy variables for different (unnamed) product types or service packages (e.g. Risselada et al. 2010, table 5). In addition, such constructs naturally need to be included in multiple studies. For example, Giudicati et al. (2013) discuss social network variables in relation to churn. While certainly relevant, this approach is relatively new, and not enough studies seem to have been published (with the above criteria) on the subject to be included in the meta-analysis.

A final yet also relatively obvious criterion is that the study needs to include output which can be transformed into an effect size, or from which effect sizes can be approximated or calculated post-hoc. It is also important to note that as I only include behavioural studies, the sample is solely restricted to correlational studies.

(31)

31

Around 7.5% of the initial set of 399 papers (30 papers) were coded, and six additional papers were included from other sources. I have set three available samples as the minimum inclusion criterion for a theoretical construct.

Figure 9. Excluded papers

(32)

32

Coding scheme and moderators

An initial version of the coding scheme was created based on earlier research. In a pre-test, this coding scheme was used to code a set of 25 papers. Based on my own experience and that of another coder, and in discussion with two advisors, the coding scheme was incrementally changed until it reached the version used for this thesis. The full coding scheme is included in appendix 4.

In this coding scheme, several types of moderators were included. My choice for these specific moderator variables originated from their inclusion in earlier, related meta-analyses (e.g. Pick & Eisend 2014), and their inclusion in unrelated meta-analyses (e.g. Bijmolt et al. 2005). I also based my choice for moderators on my knowledge of the literature, and on suggestions in other papers. For example, Assmus et al. (1984) note that systematic differences between studies can be related to four types of moderators: moderators relating to research context (i.e. situational), to model specification aspects, to measurement aspects and to estimation procedures. I summarize these categories into two categories: research methodology and situational aspects. In addition, I include one other category: publication characteristics. I will now discuss which moderator is included per category, whether I expect an effect, and the direction of such an effect, if applicable. These moderators and their expected effects are summarized in table 3 below.

Table 3. Moderators and hypotheses

Area Moderator included Expected effect

Research methodology

Model type

Stronger effect sizes for models containing only two variables (F-tests, chi square tests, t-tests) in comparison to models containing more

variables (probit, logit models)

Number of variables Weaker effect for models that contain more variables

Data source Stronger effects in survey studies in comparison to archival/combined data, due to potential reporting biases

Boundary p-value Weaker effects for models with boundary p-values Panel versus non-panel data Weaker (less inflated) effects for panel data

Situational

Region No specific hypotheses

Industry type No specific hypotheses

Contractual vs non-contractual Stronger effects in non-contractual areas, because lack of contracts allow immediate cancellation

Publication

Journal impact No specific hypotheses

Subject area No specific hypotheses

(33)

33

Moderators: Research methodology

First, I have coded model type/estimation method as a moderator. Due to differences in model assumptions, the type of model may influence the effect size in different situations. This study will use a partial effect size measure (as argued for in the next section). One of the main drawbacks of a partial effect size measure (here: the partial correlation coefficient) is that such measures ‘correct’ the correlation between two variables for the inclusion of additional variables.

Therefore, models that contain multiple variables have a higher chance of containing potentially more relevant variables, which can reduce the partial correlation coefficient between the dependent and the independent variable. Similarly, including more variables can also weaken the partial correlation between an independent and dependent variable when a third variable contains spurious correlation. Therefore, I expect that tests which by definition can only contain two variables will have stronger effect sizes than models which can and usually do include more variables. For exactly the same reason, I include the number of variables included in a model as a moderator, with the same expectations. As a third moderator in this category, I include the source of the data, which distinguishes self-reported data, archival data (i.e. company databases), and studies that use a combination of both. As individuals may have reason to misreport their behaviour, or simply may not accurately recall their behaviour, I expect that self-reported data in general contains stronger effect sizes than archival and mixed datasets. I also include a moderator that measures whether the effect size is calculated from a boundary p-value. Some studies only p-values such as p<0.05, and not report the actual p-value obtained. This may matter, as lower p-values are related to stronger effect sizes, given the same sample. Therefore, only reporting a boundary p-value may lead to effect size estimates which are too conservative. Therefore, I expect a weaker effect size in these cases.

(34)

34

Situational characteristics

In addition to aspects of research methodology, I have added several moderators which take into account situational aspects. Specifically, following the question posed in Verhoef et al. (2007), I consider whether the region where the data has been obtained influences churn estimates. There may be many reasons why this is the case. For example, regions containing larger countries may have larger markets, with more competition, which offers the possibility of higher churn rates. Similarly, a related meta-analysis has suggested that regions with a more individualistic culture are less susceptible to switching costs, and thus more likely to churn (Pick & Eisend 2014). Given the many potential reasons why regions may differ, I offer no specific hypotheses on the direction of the difference.

Again following Verhoef et al (2007), churn rates may depend on the type of industry considered. Aside from different industries containing different numbers of competitors, some industries (e.g. financial) may contain more complex products, for which comparison between alternatives is not easy. This could decrease churn rates, or change the relationship between other variables and churn rates. In addition, some industries may inherently offer more expensive products or services. It may be that consumers churn more easily when higher expenses are concerned, as the switching decision then has the potential of saving larger amounts of money. I pose no specific hypotheses per industry.

Third, I include a moderator that measures whether the study in question contains products/services that are contractual or non-contractual in nature. It should be noted that due to a limited number of studies, this is very much the same as differentiating between products and services, as almost all of the services in the primary studies are contractual in nature, while products are not. This service versus product criterion is similar to de Matos & Rossi (2008) and (Pick & Eisend 2014).

There are many potential differences between products and services and/or contractual versus non-contractual products and services. For example, services may be less easy to physically compare, which may make churn more difficult and any drivers potentially weaker. Similarly, given their contractual nature, it may also be more difficult to churn in case of contractual services, weakening the effect of any potential driver. For these reasons, I expect that studies using contractual and service data rather than non-contractual and product data report weaker effect sizes, in general.

Publication characteristics

(35)

35

significant result in comparison to other journals, which should also result in different effect sizes in journals with different impact factors.

However, there are two possible expectations: either high quality journals report stronger effect sizes due to the easier publication of significant studies, or they include weaker effect sizes due to the inclusion of methodologically ‘better’ studies which correct for heterogeneity, common method variance, or other factors that may inflate effect sizes. For these reasons, I pose no specific hypotheses. For very similar reasons, I include a moderator that measures whether the study originated from marketing journals or from other literature. While I have no reason to expect that this makes a significant difference, it may well be that studies in other types of journals tend to study other types of data, have different procedures which are considered acceptable, or simply have more or less incentive to report significant effects. Given the unknown differences between these types of journals, I do pose specific hypotheses regarding the direction or significance of the effect.

Moderators that were coded, but not included

There are several other moderators which have been coded, but which I have not used for the current analysis. For example, I have initially included a moderator for estimation procedure (e.g. OLS, 2SLS, maximum likelihood), but this is almost never mentioned in individual papers, thus cannot be used to differentiate studies. Similarly, there was very limited variation in the aggregation type (e.g. whether data was on a customer level, on a household level and so forth).

I also coded the type of firm (e.g. large, medium small, or mixed). However, this variable also contained limited variation, as the firm sizes were almost never mentioned in primary studies.8_{Another substantial}

omission was that I aimed to include a variable that measured whether variables in key categories (e.g. perceptions, behavioural/relationship variables, demographics, and so forth) were included in the primary studies in question.

This was based on a recommendations by other authors (e.g. Aloe 2015), who suggested that effect sizes from studies containing more essential variables may be more comparable to one another. In contrast, studies omitting key variables may report different correlations, above and beyond the effects of simply including more variables that I previously discussed. After all, using a partial effect size allows for such effects to occur. However, limitations in degrees of freedom have also led me to exclude this moderator from the current analysis.

8_{In addition, the simple requirement for sufficient data results in such studies usually obtaining data from large}

(36)

36

Effect size measures

Common effect size measures in meta-analyses

The ultimate aim of most meta-analyses is to quantitatively compare and generalize results across a larger body of literature. As such, any meta-analysis needs a common statistic that can be used to compare the size of the effects obtained in individual studies. There are three very common measures of so-called effect size that are used in meta-analyses (Borenstein et al. 2009, Cooper et al. 2009, Lipsey & Wilson 2001).9_{The choice for a statistic depends on the type of data under study, and on the available}

data in the studies itself. For continuous dependent variables, especially in experimental settings, it is common to compute Cohen’s d (Lipsey & Wilson 2001). This represents the standardized mean difference in the dependent variable between two groups.10_{For correlational studies, it is common to}

use the (Pearson) correlation coefficient. Finally, for binary data using 2x2 contingency tables, it is common to use the odds ratio as an effect size measure. The odds ratio is commonly transformed into the log odds ratio, as this has more desirable statistical properties.

Transformations and selective reporting

There are several difficulties in obtaining the necessary data for calculating effect sizes. Most frustratingly, it is relatively common for individual papers to report poorly on the studies performed, possibly due to space restrictions in journals, or for other reasons. For example, in several churn studies, in case of non-significant results, test statistics, p-values or other basic information are simply not reported at all (e.g. Verhoef & Donkers 2005). This makes extraction of the effect sizes difficult, usually even impossible. Moreover, as a result, not properly reporting insignificant tests means that any meta-analysis cannot take these into account. Therefore, most meta-analyses will overestimate the effect size of the variables in question. This is a very common problem (Lipsey & Wilson 2001).

In many other cases however, effect sizes can be extracted from studies, or approximated. In general, there are many different types of statistical tests, many different test statistics, and many different methods of reporting the outcomes of such tests. In the last decades, meta-analysts have made use of a number of transformations and conversion formulas which allow the (approximate) calculation of effect sizes from a wide array of different test statistics. There are several different formulas to convert between different effect size measures. For example, as shown in Hasselblad & Hedges (1995), with some assumptions, Cohen’s d can be converted to a log-odds ratio (in certain cases) by the formula:

𝐿𝑜𝑔𝑜𝑑𝑑𝑠 = 𝑑 ∗ π √3