How one question can help +

(1)

Not at

All Likely Extremely Likely

0 1 2 5 6 7 8 9 10

How likely is it that you would recommend “this thesis” to a friend or colleague?

How one question can help

The impact of recommendation intention on

customer referral and cross-buying, compared with

(2)

How one question can help

The impact of recommendation intention on

customer referral and cross-buying, compared with

overall satisfaction

Key words: Recommendation intention, Customer referral, Cross-buying, Overall satisfaction, Relationship age

Groningen, November 25th₂₀₁₀

Master thesis

Master of Business Administration

Marketing Management and Marketing Research

Department of Marketing, Faculty of Economics and Business University of Groningen The Netherlands Kasper Knol Student number: 1760947 E-mail: kasperknol@gmail.com Supervision

University of Groningen, Department of Marketing First supervisor: dr. J.T. (Jelle T.) Bouma Second supervisor: dr. J.E. (Jaap E.) Wieringa

(3)

How one question can help Management Summary

____________________________________________________________________

Management Summary

This report describes how recommendation intention affects customer referral frequency and cross-buying behaviour of existing customers and the difference in affect compared to overall satisfaction. Witch in tern provides business and science with insights how this single question metric can help map and predict customer referral and cross-buying behaviour. The research is unique as it simultaneously investigates the effect of recommendation intention and overall satisfaction, using customer level data. And is based on a single case analysis of the Dutch insurance industry.

Literature review reveals some general benefits of customer metrics, however also indicates reveals somewhat of a debate about the effect of these metrics. In general only few papers investigate the recommendation intention, in contrast to highly researched metric overall satisfaction. Based on findings of other researchers relationship age is expected to have a moderating effect on the relationship between customer referral and cross-buying. A similar effect was already found between overall satisfaction and number of products purchased in other research.

To provide a detailed analysis this research uses two different samples. One sample of 480 respondents of an online survey is used to investigate referral behaviour. The second sample consisting of 4608 customers of a large Dutch insurance company and is used to investigate cross-buying. The latter is information obtained from an internal database of a marketing research agency interlinked with the internal database of the large Dutch insurance company. The data was used to estimate three Generalized Linear Model’s (GLIM), two model based on a Poisson distribution and one normally distributed model. In addition, several other analyses and a market simulation were performed.

(4)

How one question can help Management Summary

____________________________________________________________________ that the effect of recommendation intention on customer referral and cross-buying is affected by the relationship age a customer has with a company.

(5)

How one question can help Preface

____________________________________________________________________

Preface

When I started secondary school I never imagined that I would someday be a university graduate with a double master degree in Marketing. It was a long and interesting journey, but I am very glad that I pushed myself in accomplishing this achievement.

However, I could not have done this on my own. Therefore most and foremost I would like to thank my parents for their continuous support, without which I would have never come this far. Thanks mom and dad!

In addition, I would like to thank my supervisor dr. Jelle Bouma for his support during the period I was writing this report. Also, I would like to thank dr. Jaap Wieringa for his help and feedback on my work.

Last but not least I would like to thank dr. Linda Teunter for giving me the opportunity to increase my knowledge in the field of marketing research and her help during this research.

Groningen,

(6)

How one question can help Table of Contents

____________________________________________________________________

1: Introduction

Already for many years brands use celebrity endorsers to recommend their products. For example in the 1950’s Westmore Cosmetics used Marilyn Monroe to recommend their products. For some other products it is common practice to use experts for recommendations, a well known example is using dentists in toothbrushes or toothpaste commercials. In another form, recommendations also gained importance online. Internet retailers like Amazon.com, are now able to provide real-time product recommendations to customers based on their preference, previous purchases and/or products bought by others to increase sales. However, nowadays for almost all companies it is true that they are very interested in the recommendation behaviour of their customers, and how this affects firm performance.

(9)

____________________________________________________________________ Recommendation intention is only one of many customer metrics. Customer metrics can be categorized into observable measures (involve behaviours of customers that typically relate to purchase or consumption of a product or service) and unobservable constructs (customer perceptions, attitudes and behavioural intentions) (Gupta and Zeithaml, 2006). Gupta and Zeithaml (2006) divide unobservable or perceptual customer metrics into three groups: customer satisfaction, service quality and loyalty & intentions to purchase. Recommendation intention and NPS best match with the last group. Observable measures from a customers’ perspective include decisions of when, what, how much and where to buy and from a firm’s perspective this translates into decisions about customer acquisition, retention and lifetime value (Gupta and Zeithaml, 2006). Gupta and Zeithaml (2006) suggest a framework to identify a link between what firms do (i.e. their marketing actions), what customers think (unobservable constructs), what customers do (behavioural outcomes) and how customer’s behaviour affects firm performance.

1.1: Problem statement

In order to create some more insight in the effect of recommendation intention this research will focus on observable behavioural customer metrics. Following the perspective of Gupta and Zeithaml (2006), what customers think affects what customers do and what customers do affects a firm’s financial performance. The two main observable behavioural customer metrics this study will focus on are customer referral and cross-buying behaviour.

1.1.1: Research objective

(10)

____________________________________________________________________ 1.1.2: Research questions

The main research question this report will answer is:

How does recommendation intention affect customer referral frequency and cross-buying behaviour, of existing customers? And is this effect different than that of overall satisfaction?

This research question can be divided into the following sub-questions: 1) What are the benefits of customer metrics and NPS for a company?

2) What are the effects of recommendation intention on customer referral of existing customers?

3) What are the effects of recommendation intention on cross-buying of existing customers?

4) What is the difference in effect between recommendation intention and overall satisfaction on customer referral?

5) What is the difference in effect between recommendation intention and overall satisfaction on cross-buying?

6) What should managers do based on the answers on the previous questions?

1.2: Research gap

(11)

____________________________________________________________________ intention on customer referrals and cross-buying. And at the same time include overall satisfaction in the analysis to compare the results to recommendation intention.

1.3: Scope of the research

This research will investigate one case in the insurance industry in the Netherlands. Keiningham et al. (2007) found that the best loyalty measure seems to be industry specific and that the NPS performance different between industries, for this reasons the findings of this study are not generalizable to other industries. Recommendation intention will be investigated on a micro (customer) level and not as in others studies at an aggregated level. Satisfaction scores are also included in the analysis to compare to the effects of recommendation intention. Furthermore the effects of recommendation intention on consumer behaviour will only be investigated on referral frequency and cross-buying. Other outcomes of consumer behaviour will not be investigated, this study will also not compute NPS from recommendation intention to investigate the direct link to financial performance.

1.4: Structure of master thesis

(12)

How one question can help Theoretical Framework

____________________________________________________________________

2: Theoretical Framework

This chapter contains an overview of the literature regarding the benefits of measuring recommendation intention and its relation to customer referral and cross-buying. The first section describes customer metrics and NPS, including topics as overall satisfaction, intent and loyalty. In the second paragraph customer referral will be investigated and its relationship with recommendation intention. Thirdly, cross-buying and the relationship with recommendation intention is inspected. Fourthly, the moderating role of relationship age is discussed. The chapter ends with a conceptual model in which the findings and hypotheses of the previous paragraphs are summarized.

2.1: Customer metrics and NPS

This section consists of two parts, the first part examines the general benefits of measuring customer metrics and elaborates on the customer metrics: satisfaction and intent. The second part investigates the specific benefits of NPS.

2.1.1: General benefits of Customer metrics

(13)

____________________________________________________________________ 2.1.1.1: Customer metrics and Satisfaction

The most commonly used customer perceptual metric by managers is satisfaction (Gupta and Zeithaml, 2006). It has been defined in many different words, but essentially as the consumer’s judgement that a product or service meets or falls short of expectations (Gupta and Zeithaml, 2006). Gupta and Zeithaml (2006) mention that applied marketing research tends to measure satisfaction at the transaction level but more recently as an overall evaluation, a cumulative construct that is developed over all the experiences a customer has with a firm. One definition of satisfaction that reflects the overall evaluation is the following: a persons’ overall satisfaction is an overall evaluation based on the total purchase and consumption experience with a good or service over time (Anderson et al., 1994). Consumers are asked how satisfied they are about something for example, their relationship with company X. Researchers have most focused on the following antecedents of customer satisfaction: expectations, disconfirmation of expectations, performance, affect and equity and the following outcomes of customer satisfaction: complaining behaviour, negative word of mouth and repurchase intentions (Szymaski and Henard, 2001). Several empirical studies found positive relationships between satisfaction and behavioural outcomes. For example: Bettencourt (1997) found a significant relationship between satisfaction and customer referral and Bolton (1998) between satisfaction and purchase intentions. In addition, Bolton and Lemon (1999) found a relationship between satisfaction and usage of a service, Zeithaml et al. (1996) found an effect of satisfaction on the relationship duration and Anderson et al. (2004) found a positive association between customer satisfaction and shareholder value.

(14)

____________________________________________________________________ different characteristics have different thresholds and consequently different repurchase probabilities. For example, women of more than 60 years of age whom have no children are more loyal. At the same level of rated satisfaction their repurchase rate is higher than that of other subjects.

2.1.1.2: Customer metrics and Intent

In contrast to satisfaction, asking people about there intent is more forward looking, people are asked to identify future behaviour. When it comes to predicting actual behaviour for intentions, researchers have found that statements of intentions are not always fulfilled in reality (Gupta and Zeithaml, 2006). Although Kalwani and Silk (1982) show positive correlation between intention and actual purchase behaviour, Gupta and Zeithaml (2006) argue that the predictive power of this finding is limited. Other studies find the relationship between intentions and actual behaviour is non-linear. One example of this is a study among 5000 bank customers by Kamakura et al. (2002). Customers’ likelihood to recommend has a nonlinear association with their transactions per month, number of years they stay with the bank an overall firm profits. In addition, the effects of measuring loyalty intention show that measuring intent increases the tendency for consumers to increase subsequent purchase behaviour (Morwitz et al., 1993; Dholakia and Morwitz, 2002; Dholakia and Morwitz, 2002). Research by Chandon et al. (2005) also finds that measurement of purchase intentions increases the association between latent intentions and purchase behaviour. They conclude that the effects are significant and robust across a variety of purchase behaviours, sampling frames, and ways to measure intentions and behaviour. In addition, they also show that measurement of purchase intentions does not influence purchases of consumers who have neutral purchase intentions (Chandon et al., 2005).

2.1.2: Specific benefits of NPS

(15)

____________________________________________________________________ 2.1.2.1: NPS and Loyalty

(16)

____________________________________________________________________ 2.1.2.2: NPS and other benefits

Throughout his publications (Reichheld, 2003; Reichheld, 2006a; Reichheld, 2006b) Reichheld describes multiple benefits of using NPS. One of the first is its simplicity and ease of measurement. Instead of using complex and long survey only one or two questions are enough. The results can be quickly analysed ensuring timely data instead of data that is months old (Reichheld, 2003). Because of this simple measure it is easy to communicate and motivate an organization to become more focused on improving products and services for customers. However most importantly, Reichheld (2003) concludes that a company’s NPS correlates with revenue growth and is the best predictor of growth. Figure 1 was presented in Reichheld (2003) and shows the three-year growth correlations between growth and NPS for the U.S. Airline industry between 1999 and 2002.

Figure 1: Growth and NPS for the U.S. airline industry. Source: Reichheld (2003)

Figure 1 clearly shows that airline companies with a higher NPS, especially Southwest, show a higher three-year growth. According to Reichheld (2003), the path to stainable, profitable growth begins with creating more promoters and fewer detractors and making your NPS transparent throughout your organization. Whereas Reichheld is convinced of the importance and correctness of NPS, there are sceptics whom do not totally agree with his statements.

(17)

____________________________________________________________________ growth. Based on there research they find it difficult to imagine that NPS is a superior metric. They explain that managers who are guided by the NPS may develop unrealistic views about performance, value and shareholder wealth, leading them to misallocate resources. Morgan and Rego (2006) already concluded that prescriptions to focus customer feedback systems and metrics solely on customer recommendation intentions and behaviours are misguided. However, Keiningham et al. (2007a) points out that the effectiveness of NPS on business performance cannot be accurately made from the research by Morgan and Rego (2006). According to them Morgan and Rego (2006) misunderstood the data fields from which they calculated NPS. In their own research Keiningham et al. (2007a) also fail to replicate the assertions of Reichheld (2003) that “clear superiority” of NPS compared to other measures. As for the benefit of having only one question to measure research shows that typically single item measures are less reliable than multi-item scales/constructs (Keiningham et al., 2007b). Furthermore, customers’ loyalty-based behaviours seem to be multidimensional, no one item metric best predicts all behaviours associated with customer loyalty (Keiningham et al., 2007b).

2.2: Customer referral

Customer referral of existing customers can be a means to initiate customer acquisition. Customer acquisition refers tot the first-time purchase by new or lapsed customers (Gupta and Zeithaml, 2006). Acquisition of customers can either be marketing induces or word of mouth (WOM) customer acquisition (Villanueva et al., 2008). Marketing induced acquisition is for example use of mass media and direct mail on the other hand WOM communications can be seen as newspaper articles and referrals (Villanueva et al., 2008).

Word of Mouth

(18)

____________________________________________________________________ most important factors in acquiring new customers (Jones and Sasser, 1995). Research by on a web hosting company revealed that marketing induced customers add more short-term value, but word of mouth customers add nearly twice as much long-term value to the firm (Villanueva et al., 2008).

Recommendation intention is intended to ask the WOM behaviour of existing customers and according to Reichheld (2006b) detractors are the least likely to repurchase or refer and they account for more than 80% of negative WOM, whereas promoters account for 80% to 90% of positive referrals. In addition, Blodgett et al. (1993) found that, on average, dissatisfied consumers tell about their negative experiences to nine others, and for some businesses this may lead to a loss in volume of ten to fifteen percent.

As already mentioned in section 2.1.2.1 Reichheld (2003) talks about that when customers recommend, they do more than indicate that they have received good economic value from a company, they also put their own reputing on the line. And they will only risk their reputation if they feel intense loyalty. Specific research on WOM motives by Sundraram et al. (1998) provides more insight in the reasons why people engage in recommendation/WOM. They distinguish eight different motives for WOM behaviour, four for negative and four for positive WOM. The motives for positive WOM are: altruism (the act of doing something for others without anticipating any reword in return), product involvement, shelf-enhancement and helping the company. For negative WOM they differentiate: alruism (in this case: preventing others from experiencing the same problems they did, anxienty reduction (an opportunity to vent their anger), vengeance and advice seeking.

(19)

____________________________________________________________________ find support that recommend intention is more strongly correlated to recommend behaviour than repurchase intention, customers’ perceptions of satisfaction, value and expectations (Keiningham et al., 2007b). The foregoing led to the formulation of the following hypotheses:

Hypothesis 1: Recommendation intention has a positive effect on the number of customer referrals.

Hypothesis 1a: Recommendation intention is a better predictor of number of customer referrals than overall customer satisfaction.

2.3: Cross-buying

Reichheld (2003) articulates that promoters are loyal enthusiasts who keep buying from a company. However, from this statement it is unclear if promoters keep increasing their level of products owned or keep their level of consumption the same. On the other side detractors are customers who feel badly treated by a company, so badly that they cut back on their purchases, switch to competition if they can, and warn others to stay away from the company they feel has done them wrong (Reichheld, 2006a).

(20)

____________________________________________________________________ related to customer retention, number of services used by a customer, and customer share of wallet and customer satisfaction has the biggest impact on cross-selling (Loveman, 1998). However, this finding is not generally supported by other studies. Verhoef et al. (2001) concludes, based on 2018 insurance customers whom were surveyed by telephone and portfolio changes were collected over a one year period, that there is no main affect on satisfaction on cross-buying. Cross-buying is not only affected by the degree of satisfaction but also other factors, for example customers needs and comparison offers from other suppliers (Verhoef et al., (2001). This last part was investigated by Bolton et al. (2000), they found that customer’s comparison with competing suppliers has a significant impact on service usage and retention. However, Verhoef et al. (2001) did find that as a relationship length increases the effect of satisfaction on cross-buying also increases. Consistent with the findings of Bolton (1998), based on 650 cellular phone customers over a period of 22 months. In addition, Verhoef et al. (2001) found that low satisfaction results in abandonment of services already purchased. In a later study by Verhoef and colleagues, only affective commitment shows to have a positive impact on the number of services purchased, trust and satisfaction do not seem to have an effect (Verhoef et al., 2002).

In contrast to the role of satisfaction on cross-buying, only little research is done on the effect of loyalty and intentions to purchase on cross-buying. One of the few papers on this topic is by Kamakura et al. (2002), whom find a positive path coefficient of 0.27, between customer intentions (willingness to recommend) and customer behaviour (bank share, number of transactions) based on examining bank data.

(21)

____________________________________________________________________ purchased between these two points in time. Creating that kind of dependent variable provides much more insight into the effect of cross-buying in contrast to the number of services owned used by Bolton (1998), Bolton and Lemon (1999) and Verhoef et al. (2002). This study also wants to investigate the effect of recommendation intention and overall satisfaction on real changes in the number of services owned. For that reason, the following hypotheses were formulated:

Hypothesis 2: Recommendation intention has a significant predictive power in the change of number of services owned.

Hypothesis 2a: Recommendation intention has a better predictive power in the change of number of services owned than overall customer satisfaction.

2.4: Relationship age

(22)

____________________________________________________________________ Hypothesis 3a: The effect of recommendation intention on the number of customer referrals will be different, depending on the length of the relationship age of customers with a company.

Hypothesis 3b: The effect of recommendation intention on change of number of services owned will be greater for people with a longer relationship age with a company.

2.5: Conceptual model

As described in the beginning of this chapter, the chapter will end with a conceptual model based on the literary review in the previous paragraphs and hypothesises formulated. However before that, the hypotheses formulated in this chapter will be first linked to the sub-questions formulated in Chapter one.

2.5.1: Link hypothesis and sub-questions

To clarify which hypotheses are used to answer the sub-questions formulated in Chapter one a short overview will be provide, illustrated in table 1.

Sub-questions Corresponding hypothesis

2) What are the effects of recommendation intention on customer

referral of existing customers? • H1

• H3a 3) What are the effects of recommendation intention on

cross-buying of existing customers? • H2

• H3b 4) What is the difference in effect between recommendation

intention and overall satisfaction on customer referral? • H1a 5) What is the difference in effect between recommendation

intention and overall satisfaction on cross-buying? • H2a

(23)

____________________________________________________________________ 2.5.2: The conceptual model

Based on all of the above this research proposes the following conceptual framework:

Figure 2: Conceptual model

The conceptual model displayed in figure 2 provides a short overview of the focus arias this study investigates. The relationship between recommendation intention and customer referral is investigated by looking at the number of customer referrals. In this research customer referrals is defined as the number of times customers advise other customers (e.g., friends, family, colleagues) to do business with the focal supplier, similar to the definition of Verhoef et al. (2002).

(24)

How one question can help Research Design

____________________________________________________________________

3: Research Design

This chapter describes the choice of the research and data collection methods used in this research. It will start with explaining the research methods. Secondly, the data collection process will be described. The chapter ends with a description of the analysis methods use to analyze the data.

3.1: Research method

The nature of this research will be a descriptive research to determine the effect of recommendation intention on customer referrals and cross-buying. The research will be based on a multiple cross-sectional design (Malhotra, 2007), data from more than one sample is needed to evaluate both customer referral and cross-buying. However, to assess cross-buying it is also necessary to measure some variables in two points in time, indicating a longitudinal design of only two points in time.

3.2: Data collection

To investigate the two behaviours at interest (referral & cross-buying) multiple data streams needed to be collected and combined. It was not possible to collect all data from the same sample, for that reason it is necessary to collect data from two different samples. One sample for investigating customer referrals and the other sample to investigate cross-buying. Within this process three main data streams can be distinguished: 1) primary data was collected using an online customer survey, 2) secondary data was collected from a research company database on customers’ recommendation intention scores of a large Dutch insurance company and 3) additional secondary data was collected from the same Dutch insurance company as mentioned under data stream two. This additional information consisted of data from the insurance company’s database (for the same people as the recommendation intention scores are know).

3.2.1: Online customer survey

(25)

____________________________________________________________________ Appendix A. Respondents were asked to indicate their premium insurance company and base their answers on their relationship with that insurance company.

Variable Base on Source Scale

Recommendation

intention How likely is it that you would recommend [company X] to a friend or colleague?

Reichheld

(2003) 0 not at all likely, 10 extremely likely Overall

satisfaction How satisfied are you about [company X]? Alteration of Verhoef et al. (2002) and Singh (1990)

1 very

dissatisfied, 5 very satisfied Referral Yes/No Did you really recommend

[company X] recently?

New Yes/No

Actual referral How many times did you actually recommend

[company X], since you are insured at [company X]?

New Ratio

Relationship age (in years)

How long have you been a customer of [company X]?

New Ratio

Gender What is your sex? New Nominal

Age What is your age? New Ratio

Education What is your highest education level?

New Ordinal 8-point scale

Table 2: Variables, questions, source and scales from online survey

To measure recommendation intention, the question formulation and answering scale used is similar to Reichheld (2003). Satisfaction, necessary for H1a, is measured based on an alteration of Verhoef et al. (2002) and Singh (1990) these researchers ask satisfaction of smaller constructs, for example satisfaction on personal attention (Singh, 1990) or responding to claims (Verhoef et al., 2002). Recommendation intention is also measured on a macro level (the whole company, not a specific part of the service) for this reason overall satisfaction in measured. In addition, it is asked if consumers really recommended the company and if this how many times, this is investigate hypothesis H1. The next questions measures the relationship age, necessary to test hypothesis H3a. The other questions (sex, age, education) are included to describe the sample.

(26)

____________________________________________________________________ questionnaire. In period of one week 520 respondents completed the questionnaire (response rate = 20.8%).

3.2.1.1: Data exclusion

Before the data is used for analysis a consistently check will performed on the data to identify data that are out of range, logically inconsistent, or have extreme values (Hair et al., 2006). The reasons consider for data deletion are:

• Missing values for some variables, because of the sufficient number of respondents to perform the analysis on, incomplete data will be deleted. • Unrealistic values, for example if the relationship age longer than the age of

the respondent.

• Extreme values, if the age of the respondent is higher than 100 years old or has an age below 10. Although these could be valid cases, they seem to be somewhat unrealistic to the researcher.

After checking the data on the two criteria mentioned above, 480 respondents remain for further analysis.

3.2.2: Cross-buying data selection

(27)

____________________________________________________________________ retrieving the data from the insurance company database is not equal for all individuals and ranges between 6 till 11 months. A large time period between t=0 and t=1 is preferred, to create the opportunity to observe sufficient mutations in the number of insurances.

Table 3 provides an overview of all the available data and provides some more details over the data. For each variable the source and de data/question the variable was based on is displayed, in addition the scale of the variable is provided.

Variable Based on Source Scale

Recommendation intention

How likely is it that you would recommend [company X] to a friend or colleague? Research company 0 not at all likely, 10 extremely likely Overall satisfaction

How satisfied are you about [company X]? Research company 1 very dissatisfied, 5 very satisfied Gender Research companies database Research

company 1=female 0=male Number of

services at t=0

Summation of all insurances owned at t=0

Insurance company

Ratio

Number of

services at t=1 Summation of all insurances owned at t=1 based on mutations since t=0

Insurance

company Ratio Relationship Age

(in months)

Difference between start client relationship with insurance company and t=0

Insurance company

Ratio

Age Difference between birth date client and t=0 Insurance company Ratio Time period between t=0 and t=1

Difference between date t=0

and t=1 Insurance company Ratio

Table 3: Variables, Source/question and scale of data for analysing cross-buying

3.2.2.1: Data exclusion

(28)

____________________________________________________________________ successfully linked. In the same process as paragraph 3.2.1.1 the data was examined for consistency and after checking the data 4608 people remained for further analysis. Similar to the referral data the main reasons for deleting these 18 individuals were: to many missing values and unrealistic values for some data, some respondents had an age below 10 years old and some individuals were above 100 years old.

3.3: Dependent and independent variables

Dependent variables

The dependent variable in this study to measure customer referral is a combination of the variables “Referral Yes/No” and “Actual referral”, measured in the online survey. The actual number of referrals of the participants is used combined with zero customer referrals for people that mentioned that they did not refer on the “Referral Yes/No” question.

Two dependent variables will be used for analysing cross-buying behaviour. First the variable “number of services at t=0” will be used for Model 2. In addition the dependent variable for Model 3, the change in the number of services between t=0 and t=1 is somewhat more complex. This variable is computed as follows:

Equation 1

=

!NS NS t=0 - NS t=1

where

NS = the number of services owned

t=0 = moment of measuring recommendation intention scores etc. ranges between

August 2009 and January 2010.

t=1 = June 2010, the moment the data was interlinked

Independent variables

(29)

____________________________________________________________________ Covariates

To improve the total variance explained, by Model 2 and Model 3, two covariates will be included. Model 2 (number of services at t=0) will be included with relationship age. For model 3 (difference in number of services) the variable number of services at t=0 will be included to improve the model.

3.4: Analysis method

After some general descriptive statistics of the data (mean etc.), the distributions of the dependent variables will be tested. This is done by investigating the Skewness (Malhotra, 2007 p. 462) and Kurtosis (Malhotra, 2007 p. 462) of the variables and by a Kolimogorov-Smirnov test (Malhotra, 2007 p. 485). After that scatter plots will be used to reveal patterns in the data, this will provide some general insight into the data and the relationships between the independent variables on the dependent variable. After that, associations among the variables are analysed using product moment correlation (Malhotra, 2007 p. 536), this will help to investigate the relationship between the independent variables on each other. To test the significant level (alpha) of the variables and to test the hypothesises, the threshold level is a p-value equal to or below 0.05.

(30)

____________________________________________________________________ owned between t=0 and t=1 linear regression with a normal distribution seems to be more appropriate. As the data consists of both positive and negative values, because this period people can either increase or decrease their services owned. However, to be certain of the distribution of the dependent variables this will be tested.

To illustrate the effect of recommendation intention scores even more than the output of the GLIM’s, analysis of variance (ANOVA) (Malhotra, 2007 p. 505) with a post hoc test (Hair et al., 2006 p.473) will be performed to test if the mean number of referrals significantly differs per recommendation score group. For cross-buying it will be tested if there is a significant difference between recommendation intention mean from people how increased, decreased or stay the same with there number of services purchased. In addition, the mean group values will be calculated to illustrate the differences per group.

To test H1a and see if recommendation intention is a better predictor of number of customer referrals than overall customer satisfaction scenario analyses will be used (a similar process than that of Verhoef et al., 2007), using the estimates of the Poisson regression and the original data. The recommendation intention scores will be increased with 5%, 10% and 25%, with a restriction that the score cannot exceed the maximum value of the variable of interest. The same process will be performed on overall satisfaction scores to see which variable contributes most to the number of customer referrals. This process will provide a simple illustrative example of the effect of increasing recommendation intentions scores and its effect on the number of customer referrals.

3.4.1: Model Validation

To validate the distributions of the dependent variables the will look at how the Skewness and Kurtosis differ from a standard normal distribution that has a Skewness of zero1_{and Kurtosis of three}2_{. In addition, the Kolmogorov-Smirnov test}

will test if the specified distribution holds for the variable of interest. If the significance level is lower than 0.05 the null hypothesis (that the specified

(31)

____________________________________________________________________ distribution is present) can be rejected. To test the significance of the referral model and number of services at t=0 (based on maximum likelihood) the log-likelihood function of the fitted model is compared with the log-likelihood function for the intercept only model. The deviance of a model is the difference in likelihood between the model used and the full model. The full model is the model with the maximum number of parameters that can be estimated, also known as the saturated model (Dobson and Barnett, 2008 p. 79). The difference in deviance between the fitted model and the intercept only model follows a Chi-square distribution with a specific amount of degrees of freedom. This is the difference between the degrees of freedom of the fitted model and the intercept only model. Based on these results it can be tested if the model is a significant improvement to the intercept only model. To test the goodness of fit a pseudo R-squared statistic (Dobson and Barnett, 2008 p. 137) will be calculated. This will assess the amount of variance explained of the dependent variable that is explained by the independent variables in the model. However, the sampling distribution of pseudo R-squared is not easily determined (so p-values cannot be obtained), and it increases as more parameters are added to the model (Dobson and Barnett, 2008 p. 137). To assess the overall fit of the difference in number of services model and the significance, the R-square (Hair et al., 2008 p. 209) and, by means of an ANOVA (Hair et al., 2008 p. 440), the F ratio (Hair et al., 2008 p. 192) of this model is calculated.

To see whether the variables in the model contribute to the prediction of the dependent variable the coefficient estimates are divided by the standard error of the estimate to obtain the t-ratio. Following that distribution the p-value is calculated. For the Poisson models the p-value are calculated based upon the z-value. In both cases to ensure statistical significance of the model and the regression coefficients, the model and coefficients will be tested based on the statistical significance level of 5%.

3.4.1.1: Disturbance term assumptions

(32)

____________________________________________________________________ Multicollinearity: A state of very high intercorrelations among independent variables (Malhotra, 2007 p. 561). Since both models are expected to contain some interaction effects, problems with multicollinearity (Leeflang et al., 2000 p. 347) are to be expected. For this reason Variance Inflation Factors (VIF) (Hair et al., 2008 p. 201) scores will be calculated. A high value on the VIF indicates a high degree of multicollinearity, a threshold level of 10 considered to judge if the model suffers from multicollinearity. One of the possible remedies for multicollinearity is use mean-centering (subtracting the mean from the original value) the variables influenced by multicollinearity.

Non-normality: The disturbances need to be normally distributed for the standard test statistics for hypothesis testing and confidence intervals to be applicable (Leeflang et al., 2000 p. 343). If normality is not satisfied a wrong impression of the p-value is given. Therefore the normality of the disturbances is tested with two common tests, the Kolmogorov-Smirnov test and the Shapiro-Wilk W-test (Leeflang et al., 2000 p.344) and in addition with normal probability plots (Dobson and Barnett, 2008 p. 35). If non-normality seems to be present in the model the consequences of its effects will be judges on the following two aspects:

• The sensibility of the parameter estimates. Are the estimates relatively the same for the models?

• The p-values. The t-test is relatively not sensitive. Therefore highly significant outcome are expected to be valid outcomes, despite the non-normality issues. Interpretation of borderline values (close to 0.05) may be wrongly because of the non-normality of the disturbances and therefore interpretation of these values should be done cautiously.

(33)

How one question can help Results

____________________________________________________________________

4: Results

This chapter will describe the relevant results of the research performed. First, the descriptive statistics and insights of the variables of the customer referral sample will displayed. Secondly, the result of the customer referral model presented and hypothesises formulated in Chapter two (on customer referral) will be tested, followed by a simulation aimed to tested H1a. Thirdly, the descriptive statistics and insights of the cross-buying variables will be presented. The chapter ends with presenting the results of two models that describe cross-buying and more specifically test hypothesis H2, H2a and H3b.

4.1: Descriptive statistics sample referral

Table 4 presents the (mean) values of the referral data. Among other things it shows that the customers are rather pleased witch their premium insurance company, indicated by the high recommendation intention score and overall satisfaction score. Also, the sample consist of a wide range of different customers based on the demographic variables indication a good representation of the whole market. In addition to the information displayed in table 4, 59.2% of the 480 respondents did not recommend their premium insurance since they became a client. Also, detractors account for 26% of the total sample, passives for 59% and promoters for 15%. Resulting in a NPS score for the total sample of -11.

Variable Values

Number of customer referrals Mean score 1.67

Recommendation intention Mean score 7.05

Overall satisfaction Mean score 3.98

Relationship age (in years) Mean relationship age 13.05 years

Gender Women

Men

52.1% 47.9%

Age Mean age

Ranges between 14 and 77 years

43.81 years

Education Lower level learning

Lower business learning Secondary general education Secondary vocational education Higher general education

(34)

____________________________________________________________________ Investigating the distribution of the dependent and independent variables show that the Skewness and Kurtosis values differ from the expected ranges from a standard normal distribution. For example the Kurtosis of the number of customer referrals is 18.4. Full details of the Skewness and Kurtosis values, histogram and Kolmogorov-Smirnov tests can be found in Appendix B. Also, the histogram of the variable does not seem have a nice bell shape that is expected from a normal distribution. Instead it has a decreasing shape with a lot of observations on the left side and fewer if you go to the right. A shape similar to what would be expected from a Poisson distribution. In addition, Kolmogorov-Smirnov tests for a normal distribution on the number of customer referrals has a significance value of 0.000, indicating that the data is not expected to follow a normal distribution. A Kolmogorov-Smirnov test for a Poisson distribution and exponential distribution also has a significance value of 0.000, indicating that the distribution of the dependent variable according to this test does not follow these distributions. However, as the shape of the histogram somewhat looks like a Poisson distribution and the data consists of count data the model will be estimating a model based on a Poisson distribution of the dependent variable. In addition, a test model estimated based on normal distribution of this dependant variable explained substantially less variance and provided insignificant results compared to Model 1.

To gain some knowledge on the effect of the independent variables on the dependent variables before estimating a model, scatter plots were calculated. The scatter plots can be found in Appendix C. The scatter plots of recommendation intentions and overall satisfaction on the number of customer referrals show a pattern of high score of either of these independent variables increases with higher the number of referrals. Subsequently low scores correspond with few referrals. No specific pattern is visible between relationship age and number of referrals. Although the scatter plot of gender provides some indication that men refer more often than women, this difference is not significant (p = 0.100) based on an ANOVA analysis (Appendix C).

(35)

____________________________________________________________________

Means, Standard Deviations and Pearson correlations (n=480)

Mean

Standard

Deviation Y1 X1 X2 X3

Number of customer referrals (Y1) 1.67 3.23 1

Recommendation intention (X1) 7.05 1.79 0.28** 1

Overall satisfaction (X2) 3.98 0.70 0.26** 0.65** 1 Relationship Age (X3) (in years) 13.05 11.35 0.15** 0.04 0.08 1 **. Correlation is significant at the 0.01 level (2-tailed).

Table 5: Correlation matrix cross-buying

The correlation between recommendation intention on number of customer referrals is the largest (0.28) for any independent on dependent relationship. The correlation coefficient between recommendation intention and overall satisfaction, both independent variables, is 0.65. This may result in some multicollinearity problems as relatively high correlations indicate multicollinearity problems (Leeflang et al. 2000).

4.2: The Models – Customer referral

This part will first show the estimated model for the number of customer referrals. Secondly, the result of a simulation with increased recommendation intention scores and overall satisfactions scores will be presented.

4.2.1: Model 1

(36)

____________________________________________________________________

Model 1: Number of customer referrals

glm (formula = Referrals ~ Recommendation intention + Overall satisfaction + Relationship age + Recommendation intention * Relationship age, family =

Poisson(log) Coefficients: Estimate

Std.

Error z-value p-value VIF (Intercept) -1.256 0.314 -4.005 <0.0001 Recommendation intention 0.367 0.037 9.962 <0.0001 1.766 Overall satisfaction 0.370 0.077 4.828 <0.0001 1.614 Relationship age 0.024 0.003 7.569 <0.0001 1.353 Recommendation intention * Relationship age -0.008 0.002 -4.364 <0.0001 1.488

Null deviance: 1947.1 479 degrees of freedom Residual deviance: 1564.1 475 degrees of freedom

McFadden's Pseudo

R-Square 0.197 Significance model

(p-value)

<0.0001

Table 6: Model 1: Number of customer referrals (p-values calculated based on the z value using http://www.graphpad.com/quickcalcs/pvalue1.cfm)

(37)

____________________________________________________________________ evidence of the effect of recommendation intention on customer referral, therefore H1 is accepted. In addition, the data shows that detractors account for 12.8% of the WOM behaviour, passives 50.2% and promoters for 37%.

Figure 3: Mean number of customer referrals split by recommendation scores (RI)

Continuing with the interpretation of Model 1, overall satisfaction does also have a highly significant (p = <0.0001) impact on the number of customer referrals. The corresponding estimate for this variable is 0.370. The estimate is slightly higher than that of recommendation intention (E = 0.367), however recommendation intention is measured on an 11-point scale were as overall satisfaction on a 5-point scale. Taken this scale difference into account, it seems that more customer referrals can be generated by recommendation intention. This provides some indication that recommendation intention is a better predictor than overall satisfaction in predicting the number of customer referrals. However, before accepting H1a more analysis will be conducted into this relationship. The results of this will be provided in paragraph 4.2.2: Simulations.

Relationship age also contributes significantly to the model (p = <0.0001) and has a positive estimate of 0.024. In addition, the interaction effect between recommendation intention and relationship age is also highly significant (p =

(38)

____________________________________________________________________ customer has with their premium insurance company. The negative sign implies that the effect of the two interaction variables is corrected downwards. This relationship is illustrated in figure 4, based on the estimates of Model 1 and the mean and standard deviations of the variables recommendation intention and relationship age. This example is calculated based on distinguished low recommendation intention (RI) and low relationship age as the mean value of this variable minus the standard deviation. High recommendation intention and high relationship age are distinguished as the mean value of this variable plus the standard deviation. Resulting in the following values: Low RI = 5.26, High RI = 8.84, Low Relationship Age = 1.7 years and High Relationship Age = 24.4 years. This example shows that the difference between high recommendation intention scores at a high and low relationship age is relatively smaller than that the difference for low recommendation intention scores at a high and low relationship age. Hence, there is support for H3a.

Figure 4: Interaction effect recommendation intention (RI) and relationship age

(39)

____________________________________________________________________ 4.2.2: Simulations

A simulation analysis using the estimates of Model 1 was performed to investigate H1a even more beyond the finding of paragraph 4.2.1. Figure 5 illustrates the results of this simulation. Increasing the recommendation intention scores and keeping the other variables the same with 5 %, results in a 3.5% increase in number of customer referrals. Increasing scores with 10% results in a 7.1% increase and 25% in an increase of 16.6%. The same process performed on overall satisfaction results in a smaller increase in percentage of customer referrals: 5% increase > 2.8%, 10% increase > 5.6% and 25% increase > 14.0%. Although the difference between the two variables is small, recommendation intention seems to contribute more to the number of customer referrals.

Figure 5: Simulation analysis recommendation intention and overall satisfaction (Restriction max value RI=10 & OS=5)

(40)

____________________________________________________________________

4.3: Descriptive statistics sample cross-buying

Table 7 presents the (mean) values of the referral data based on 4608 participants.

Variable Values

Recommendation intention Mean score 7.77

Overall satisfaction Mean score 4.06

Number of services at t=0 Mean score 3.98

Number of services at t=1 Mean score 3.95

Time period between t=0 and

t=1 Mean age Ranges between 6 and 11 months 8.81

Relationship age (in months) Mean relationship age 166.32 months (13.9 years)

Gender Women

Men

29.2% 70.8%

Age Mean age

Ranges between 19 and 99 years

55 years

Table 7: Values cross-buying variables

The customers seem to be rather satisfied with their insurance company and the customers participated in this research have similar demographics than the profile of the whole insurance company database. In addition, to the information presented in table 7 detractors account for 13.1% of the total sample, passives for 57.9% and promoters for 29%. Calculating the NPS score of the total sample indicates a NPS of 15.9. The absolute mutation in services between t=0 and t=1 ranges from -12 to +7.

(41)

____________________________________________________________________ Kolmogorov-Smirnov test for an exponential distribution excludes also has a significance value of 0.000 and a lot of values are excluded do to that these values are outside of the specified distribution range. Although, the results do not exactly reveal the distribution of the variables, the models will be estimated based on the distribution, discussed and in paragraph 3.4.

To gain some knowledge on the effect of the independent variables on the dependent variables before calculating a model scatter plots were calculated. The scatter plots can be found in Appendix G. The scatter plots of the independent variables overall satisfaction, recommendation intention and relationship age do not indicate a specific pattern in their relationship with difference in number of services. However, the number of services at t=0 shows a somewhat negative linear relationship with difference in number of services. People who have only few insurances to start with increase there services more often and people who already have a lot of services seem to be more likely to decrees the amount of services.

As in the paragraph 4.1, to gain insight into the associations among independent variables, the correlation coefficients, displayed in table 8, were calculated.

Means, Standard Deviations and Pearson correlations (n=4608)

Mean St. Deviatio n Y1 X1 X2 X3 X4 , Y2 Difference in number of services (Y1) -.04 1.180 1 Recommendation intention (X1) 7.77 1.608 .093** 1 Overall satisfaction (X2) 4.06 0.889 .048** .317** 1 Relationship Age (in months) (X3) 166.3 6 117.438 -.038** .071** .048** 1 Number of services at t=0 (X4) (Y2) 3.98 2.494 -.386** .072** .048** .248** 1 **. Correlation is significant at the 0.01 level (2-tailed).

*. Correlation is significant at the 0.05 level (2-tailed).

(42)

____________________________________________________________________ The number of services at t=0 has the highest correlation (-0.386) with the dependent variable (Y1). This highly significant correlation is negative, corresponding with the findings from the scatter plot. The correlation between the independent variables is rather low with the highest correlation coefficients of 0.32, between recommendation intention and overall satisfaction.

4.4: The Models – Cross-buying

This part will provide the results for the models estimated to investigate cross-buying.

4.4.1: Model 2

Before we calculate the models to answer the hypotheses, we will first estimate a simple model of all the original independent variables on the number of services at t=0. This is to investigate if the customer metrics measured in t=0 also give an indication of the number of services already owned at t=0. This model is displayed in table 9.

Model 2: Number of services at t=0

glm (formula = Number of services at t=0 ~ Recommendation intention + Overall

satisfaction + Relationship Age, family = Poisson(log) Coefficients: Estimate

Std.

Error z-value p-value VIF

(Intercept) 0.955 0.005 20.723 <0.0001 Recommendation intention 0.019 0.0004 3.932 <0.0001 1.107 Overall satisfaction 0.015 0.0008 1.757 0.0789 1.105 Relationship age 0.001 0.000006 20.491 <0.0001 1.005

Null deviance: 6819.7 4607 degrees of freedom Residual deviance: 6364.2 4604 degrees of freedom

McFadden's Pseudo R-Square 0.067 Significance model (p-value) <0.0001

(43)

____________________________________________________________________ The difference between the deviance of the intercept only model (6819.7) and the estimated mode (6364.2) is 455.5 with 3 degrees of freedom (4607 - 4604). This results in a p-value of <0.00014 indicating that the model is a significant improvement compared to the intercept only model. The total amount of variance explained by the model was calculated using McFadden's Pseudo R-Square and corresponds to 6.7% of the total variance. Interesting finding is that overall satisfaction, a somewhat more back looking customer metric, does not have a significant relationship (p = 0.0789) with the number of services owned at that point in time. However, recommendation intention with an estimated beta of 0.019 does have a highly significant effect (p = <0.0001). It seems that recommendation intention, a somewhat more forward looking customer metric, also is an indication of earlier behaviour. Relationship also has an effect on the number of services people own at t=0 with an estimate of 0.001 meaning that the longer a person is a client of a company the more services he or she owns at t=0.

4.4.2: Model 3

Model 3 shown in table 10 investigates the effects on difference in number of services and will be used to test hypothesises H2, H2a and H3b. The model explains 16.9% of the variance (R Square = 0.169) and is highly significant (F ratio = 186.57, p = 0.000). However, based on the output of the Kolmogorov-Smirnov (p = 0.000) & Shapiro-Wilk (p = 0.000) tests and Q-Q plot the residuals do not seem to be normally distributed as the dots are not on or closely to the straight line, the full output can be found in Appendix H. Indicating non-normality in the model. This has, as mentioned in paragraph 3.4.1.1, consequences for the interpretation of the p-values.

(44)

____________________________________________________________________

Model 3: Difference in number of services

Difference in number of services ~ Recommendation intention + Overall satisfaction + Relationship age + Recommendation intention * Relationship age + Number of

services at t=0 Unstandardised Coefficients Standardis ed Coefficien ts Collin earity Statist ics β Std.

Error Beta t Sig. VIF

(Constant) 0.565 0.082 6.915 0.003 Recommendation intention 0.081 0.010 0.111 7.781 0.000 1.120 Overall satisfaction 0.041 0.019 0.031 2.17 0.030 1.113 Relationship age -0.001 <0.000 0.054 3.885 0.000 1.069 Recommendation Intention * Relationship age <0.000 <0.000 0.030 2.246 0.025 1.003 Number of services at t=0 -0.193 0.007 -0.409 -0.409 0.000 1.069 Model Summary R R Square Adjusted R Square Std. E. of the Estimate 0.411 0.169 0.168 1.077 ANOVA Sum of Squares df Mean Square F Sig. Regression 1081.196 5 216.24 186.57 0.000 Residual 5333.752 3602 1.159 Total 6414.948 3607

Table 10: Model 3: Difference in number of services

(45)

____________________________________________________________________ higher recommendation intention score at t=0. Thus, it appears that recommendation affects the difference in number of services, and therefore H2 can be accepted.

Figure 6: Mean recommendation intention split by change group

Overall satisfaction also has a significant relationship (p = 0.030) with the dependent variable. However, do to the normality issue in the model the impression of the p-value may be wrong, especially as the p-p-value is close to the borderline (p = 0.05). Comparing the standardised beta of recommendation intention and overall satisfaction, and thereby accounting for scale differences, recommendation intention has a much higher standardised beta of 0.111 compared to overall satisfaction (st. β = 0.031). This indicates that recommendation intention is a more important variable in explaining the variance in the model and as the significance of overall satisfaction is questionable (do to the normality issue), H2a is accepted.

(46)

____________________________________________________________________ is illustrated in figure 7, based on the estimates of Model 3 and the mean and standard deviations of the variables recommendation intention and relationship age. Similar to interaction example for the number of customer referrals, low or high recommendation intention and relationship age is distinguished as the mean value of the variable minus or plus the standard deviation of that variable. Resulting in the following values used in this example: Low RI = 6.16, High RI = 9.39, Low Relationship Age = 48.90 months (approximately 4 years) and High Relationship Age = 283.80 months (approximately 24 years).

Figure 7: Interaction effect recommendation intention (RI) and relationship age

The VIF scores for relationship age (VIF = 23.1) and Recommendation Intention * Relationship age (VIF = 26.9) were very high, indicating multicollinearity. Also in this model the variables recommendation intention and relationship age were mean-centred. The result, as can be seen in table 10, VIF values around 1 indicating no more multicollinearity problems in Model 3.