PROFESSIONAL FOOTBALL FAN TYPOLOGIES: A LATENT CLASS ANALYSIS.

(1)

PROFESSIONAL FOOTBALL FAN TYPOLOGIES: A

LATENT CLASS ANALYSIS.

Thesis MSc. Marketing Intelligence Faculty of Economics & Business

University of Groningen Broerstraat 5 9700 AB Groningen, NL

050 - 363 9111 By: Kasper H. Kuipers k.h.kuipers.2@student.rug.nl

+31 631932249 S2183544

Supervisor: prof. dr. T.H.A. Bijmolt Supervisor: prof. dr. R.H. Koning

Supervision FC Groningen: E. Froma

A thesis submitted in partial fulfilment of the requirements of the University of Groningen for the MSc. degree of Marketing 2016-2017*

[ June 28, 2017 ]

Word count: 12.013 Page count: 45

(2)

Professional Football Fan Typologies: A Latent Class Analysis.

(3)

M.Sc. Thesis MI - University of Groningen, 2017

3

PROFESSIONAL FOOTBALL FAN TYPOLOGIES: A

LATENT CLASS ANALYSIS

Abstract

To structure the heterogeneity among committed supporters of a professional football club, a LC segmentation framework is proposed that identifies homogeneous segments within that supporter base. This thesis provides insights in similarities and differences between sport fans’ preferences, supported by the combined outcomes of both a traditional hierarchical clustering and a latent class analysis. Traditionally, cluster analysis provides insights in psychographics and demographics, whereas in the past decades, latent class (LC) analysis has become a widely-used technique in market research (Vermunt, 2003). This latter is modeled with a choice-based conjoint analysis. After validation, the identified profiles in the customer database of FC Groningen are the Outsiders, Social-supporters, Joe-public, Fanatic-citizens, Price-sensitives, Committed-critics, Seat-potatoes, Committed-socials and the Ambience-seekers. Hereby, I agree with the previous literature in this field and add some new insights by providing an overarching framework.

JEL Classification: M31 • Z2.

(4)

4

PROFESSIONAL FOOTBALL FAN TYPOLOGIES: A

LATENT CLASS ANALYSIS

Preface

(5)

5 Table of Content

1. Introduction ... 6

1.1. Relevancy and Contribution ... 7

1.2. Aim of the Research ... 7

1.3. Research Methods: ... 8

2. Theoretical Background ... 9

2.1. Market Segmentation ... 9

2.2. Segmentation Bases & Methods ... 10

2.2.1. Traditional Segmentation Methods ... 12

2.2.2. Benefit Segmentation Method ... 12

2.3. Sport Fan Segmentation ... 12

2.4. Research Model ... 16

3. Methodology & Empirical Framework ... 18

3.1. FC Groningen ... 18

3.2. Study (1) ... 18

3.2.1 Design Study (1): Hierarchical Clustering Survey ... 18

3.2.2. Segmentation Method ... 19

3.2.3. Model Specification Study (1) ... 20

3.2.4. Sample Characteristics Dataset (1) ... 20

3.3. Study (2) ... 21

3.3.1. Design Study (2): Benefit Component Survey ... 21

3.3.2. De-compositional Choice-Based Conjoint Analysis ... 22

3.3.3. Model Specification Study (2) ... 23

3.3.4. Classification ... 24

3.3.5. Validation: Assessment and Comparison of Fit ... 25

3.3.6. Sample Characteristics Dataset (2) ... 25

4. Estimation Results Empirical Analysis ... 27

4.1. Discussing the Results ... 27

4.2. Study (1): Hierarchical Clustering Method ... 27

4.3. Study (2): Benefit Segmentation Method: Conjoint Analysis ... 29

4.3.1. Relative Importance ... 32

4.3.2. Profiles study (2) ... 32

4.4. Segmentation Framework Study (1) & (2) ... 33

4.5. Managerial Implications / Strategy ... 36

5. Conclusion ... 37

5.1. Concluding Remarks ... 37

5.2. Limitations & Future Research ... 38

6. References ... 39

7. Appendices ... 41

7.1. Variable Definitions... 41

7.2. Academic foundation attributes used in the Conjoint study (2)... 42

7.3. Determining the number of clusters “Clustering method” ... 43

(6)

6 Professional Football Fan Typologies: A Latent Class Analysis.

1. Introduction

This study focuses on organizations within the field of sport management. In today’s heated competition for financial resources, sport entities must perform more efficient than ever and turn more and more towards their own spectator market (Wengler, Chatrath, & Berlin, 2009). Sport consumers are a heterogeneous population (Tokuyama & Greenwell, 2011; Weed & Bull, 2004) and increasing revenues on a match day requires a consumer focus (Houston, 1986). Marketing in industrialized countries demands responding to the heterogeneity of customers (Gankema & Wedel, 1992). This implies a better understanding of the consumer’s needs, who have varying wishes and requirements (Woratschek & Beier, 2001). A useful concept in developing an effective marketing strategy is consumer segmentation (Wagner, Kamakura, & Russel, 1989). It is the subdivision of a heterogeneous group of consumers into smaller, but homogeneous groups of consumers sharing similar characteristics (Wedel & Kamakura, 2000). Market segmentation is a hot topic for organizations looking to optimize their strategy. The challenging debates regarding segmentation criteria (how), segments-based needs (what), and segment prioritization (approach) are rapidly evolving and these consumer insights are interesting for all sorts of organizations.

“The idea of dividing a market up into homogeneous segments and targeting each with a distinct

product and/or message, is now at the heart of marketing theory”

– Michael J. Croft, 1994 –

“...most entities have moved from mass marketing to segmented marketing, in which they target

carefully chosen submarkets or even individual consumers”

– Kotler and Armstrong, 2009 –

(7)

7

1.1. Relevancy and Contribution

The sports industry is one of the ten biggest business sectors. It generates its main source of revenue from its fans (DeSarbo & Madrigal, 2011). Especially the football sport is highly commercialized. The application of market segmentation in sports management increased just over the last 15-20 years (Smith & Stewart, 1999), and only few studies were conducted to identify segments for the sports industry in general (Woolf, 2008). The fans of a football club can be considered the customers of an organization and should be treated as such. Understanding motivation of sport consumers is complex (Funk & Bruun, 2007). As a store manager needs to know, what drives their customers in making an online or in-store purchase, management needs to know, what drives their fans to support their club and attend match days. These insights can lead to useful strategies and, ultimately, increased match day revenues, e.g. ticket and merchandise sales. Fan classification is of main importance for football clubs. Each fan type requires different methods of targeting and motivation. Some people might be extremely intrinsically motivated by the sport and performance itself, while others are only seeking social acceptance when deciding to attend a match day. Finding a balance between these methods is a challenging task, but crucial to gain customer insights and creating both competitive, as well as customer value (Samra & Wos, 2014).

Whereas literature indicates the benefits of a good segmentation strategy (Haley, 1968), the implication in the sports industry is still marginal. Factors explaining this phenomenon might be the great differences, and thus complexity, within and among the fan base, and the somewhat prevailing traditional football culture. However, it is these varying spectators’ requirements and wishes that make this market so interesting to understand (Woratschek & Beier, 2001), and to support the entities’ effort, market segmentation is a powerful tool (Wedel & Kamakura, 2000).

1.2. Aim of the Research

(8)

8

Furthermore, I defined the following measures to reach these goals:

(1) Construct valid and reliable questionnaires/survey’s to;

(2) Measure preferences within a large heterogeneous fan base and;

(3) Identify and measure homogeneous groups of fans with similar characteristics;

Ultimately, although not the purpose of this research study, the different segments (e.g. most price-sensitive or highest growth potential) can be:

(4) Targeted with different strategies.

In a unique cooperation with FC Groningen, a regional professional Dutch football club of the Eredivisie, I apply market segmentation by assessing the most appropriate criteria to segment their supporters on. The results are expected to be generalizable to other regional football sport entities in the Netherlands and give in-depth insights into the Dutch sport-spectators market. Ultimately, the club should gain information on how to (re)design its product-service offerings and targeting strategies to increase the attractiveness, attract existing and new consumers to come to the stadium more often and increase match day revenues (Wengler, Chatrath, & Berlin, 2009).

1.3. Research Methods:

Segmentation criteria researched before are mostly defined as psychographic [identification, attitude, loyalty, involvement] and/or demographic [gender, age, income]. In this study, I too perform an analysis to see if the supporters show similarities or differences on certain of these characteristics. First, supporters are classified according to their commitment with the club, personality characteristics and behavior. By performing a hierarchical cluster analysis, the supporters are segmented based on these demo- and psychographics. Next, they are compared on their similarities and differences. Furthermore, they are visualized on the dimensions of Age and Commitment. A survey to collect the data is distributed via FC Groningen’s website, which can be accessed through supporters’ club accounts. A limitation in these psychographic studies is, that the criteria are of a somewhat descriptive nature and provide poor predictions of future behavior, as these factors aren’t causal. As lifestyle values are interesting to know, it is also interesting to identify the actual consumer requirements. Therefore, more adequate criteria for market segmentation should include variables that drive the consumers buying behavior (Assael & Roscoe, 1976).

(9)

9

parameters differ across groups/clusters. This is modeled with a choice-based conjoint analysis, where a mixed multinomial logit model is used to estimate the parameters. Supporters are segmented on actual needs/preferences, rather than on a list of questions describing their behavior and/or demographics. They choose between actual choice sets, to pursue a ‘benefit segmentation’. and predict actual behavior. Ten product-service offerings (each consisting of six attributes and three/four levels) to be evaluated, form the basis for this benefit segmentation. Following this procedure, the fans’ behavior is more of a predictive nature and the likelihood of a future action can be estimated better. After applying the limit conjoint analysis, I compare the segmenting results and see if some similarities or differences can be observed. Finally, the identified (demographic) customer characteristics can be used to describe and profile each of the segments.

These goals are achieved with the support of multiple statistical techniques. A benefit component survey is constructed with my preference-lab and distributed via the season card holder newsletter [12.000 receivers]. Furthermore, the respondents’ data are empirically analyzed with the programs Latent Gold (Statistical innovations, version 5.1) and R (Rstudio, version 1.0.143). Hereby, I significantly contribute empirically and conceptually to the sport marketing literature as this combination of segmentation methods is not published before in this field of study.

The remainder of this paper is structured as follows. Section II reviews the available literature concerning the topic and the conceptual model. Section III describes the methodology. IV contains the empirical analysis, while section V concludes with ending remarks and recommendations.

2. Theoretical Background

2.1. Market Segmentation

(10)

10

‘Pursuing effective marketing requires an organization to respond to the heterogeneity of their customers. Segmentation has thereby become one of the most fundamental concepts in developing an effective marketing strategy (Gankema & Wedel, 1992). Both the choice of the method and the segmentation criteria influence the identification of segments. The selection of the appropriate criteria and methods is therefore of main importance for a correct segmentation (Gankema & Wedel, 1992). The different segmentation bases and methods, as well as their strengths and weaknesses, are discussed further on in this chapter.

2.2. Segmentation Bases & Methods

An effective categorization of customers in segments strongly depends on the segmentation bases and segmentation methods used. Below, I will discuss the different variables and bases that are suitable for segmentation, different segmentation methods, and their effectivity in practice.

A segmentation bases is a group of customer characteristics, used for segmentation. These different bases are general characteristics (independent from the product, service or customer situations) and product-specific characteristics (related to the product, service or specific situations). Furthermore, these characteristics can be measured directly or indirectly (Frank, Massy, & Wind, 1972). Gankema & Wedel (1992) propose six criteria for an effective segmentation. These are: identifiability, accessibility, size, stability, edibility, and homogeneity in response. Combining the four segmentation bases, and comparing them on the criteria of segmentation, leads to the following classification framework:

➢ Direct observable, General segmentation bases

These bases comprise of cultural (e.g. nationality/religion), geographic (region/urbanization), demographic (age/gender/family) and sociographic variables (income/education) used for segmentation. Segments that are based on these variables have a good identifiability, accessibility and stability (Frank, Massy, & Wind, 1972). However, the authors note that the link between these variables and actual buying behavior is often weak. Therefore, for my research these bases are not the most effective to arrive at an effective segmentation solution. They are however very useful for describing the segments afterwards, as passive variables.

➢ Direct observable, Product-specific segmentation bases

(11)

11

variables, and therefore accessibility of these segments is often very limited (Frank, Massy, & Wind, 1972). They point out that buyers from different segments, based on differences in brand loyalty don’t react very differently on marketing instruments. A small difference in the other variables show moderate differences in response, proving the usefulness of these bases. However, the accessibility stays very problematic. Furthermore, real indications on how to fill in the marketing mix (campaigns and price strategy) remain unclear from these data (Gankema & Wedel, 1992).

➢ Derived, General segmentation bases

These bases include variables regarding psychographics, personality characteristics, and lifestyle values. These characteristics are of an individual psychologic nature. The connection between personality and buying behavior is not so strong according to the empirical research of Frank, Massy & Wind (1972). Lifestyle is measured by activities, interests and opinions of customers (Lazer, 1963); (Wells, 1975); (Plummer, 1974). As argued by Wells (1975) and Dickson (1982), lifestyle is not able to explain the specific behavior regarding a specific product or brand, as the direct observable general variables. They are however very valuable in giving a colorful representation of the market and bring typologies to life (Wedel & Kamakura, 2000). The segments that are found have a good measurability, accessibility and give indications for marketing strategies. They have a limited homogeneity, because the missing connection with buying behavior. Moreover, there is limited evidence about the stability of these segments.

➢ Derived, Product-specific segmentation bases

(12)

12

2.2.1. Traditional Segmentation Methods

Traditionally, consumers are segmented on the use of personal characteristics including demographical, socio-economic and psychographic data. The former two are describing the social environment of an individual, while the latter is concerned with classifying people according to their personality traits, opinions, interests (Vermunt, 2003) and attitudes (Ahmad, 2003). The demographical and socio-economic (direct observable, general) data are useful in identifying the size, judging the attractiveness, and accessing the different segments, while the psychographic (derived, general) data is more useful for predicting buying behavior and consumption. A limitation however, is that data concerning opinions and interests tend to change considerably over time, implying that the process of segmentation should be performed continuously. Also, psychographic variables do not have a strong explanatory power. According to a study of Kassarjian (1971), general personality variables can explain no more than 10% of behavioral differences, such as in e.g. choosing brands. Moreover, since there is a wide and diverse variability of these characteristics, it is impractical to try to capture all of them.

2.2.2. Benefit Segmentation Method

Benefit segmentation is a technique where consumers are segmented based on desired benefits (derived, product-specific). These benefits are presented as attributes of a market offering (product/service) and cause consumers to purchase those, rather than merely describing them in terms of psychographic, demographical or socio-economic characteristics (Ahmad, 2003). The main advantage of this method over more traditional methods is that it gives insights in the reasons behind consumers buying or preferring certain products/services. Managers are better able to develop and optimize market offerings, in terms of product/service attributes, by knowing what benefits their consumers and what they are willing to use and pay for those beneficial offerings.

2.3. Sport Fan Segmentation

(13)

13

Mahony, Madrigal and Howard (2000) developed a psychological commitment to team (PCT) scale to segment sport consumers based on loyalty. They divide sport consumers on multiple dimensions. Namely, a psychological commitment and a behavioral consistency dimension to determine one’s attitude loyalty or the strength of commitment to a sports team. They found fans to be truly loyal, spuriously loyal, latent loyal or low loyal when supporting their team.

Furthermore, Samra & Wos (2014); Smith & Stewart (1999); Wann & Branscombe (1993), Mullin (1993), Kahle (1996) and Sutton et al. (1997) also studied the segmentation of sport consumers, based on psychographic variables. Samra & Wos state that ‘fans possess a strong and intense emotional attachment with the consumption objects’. Grossberg (1995) observed that a fan is more closely associated with a form of intensity or affect compared to general customers. Quick (2000) reviewed several sport fan studies and suggested that not all fans are motivated by the same factors. The conceptual review by Samra & Wos differentiated fans between motivation, commitment to the brand, personal commitment, enduring involvement and situational involvement. They concluded with an overarching framework classifying each fan as: ‘temporary, devoted or fanatical’. After studying different perceptions about sport, Smith & Stewart (1999) found the ‘Passionate partisans, Champ followers, Reclusive partisans, Theatregoers & the Aficionados”. Rein et al (2008) mentioned the usefulness of combining demographics with psychographic variables. First, they differentiated individuals based on demographic characteristics and then they looked for attitudinal similarities within those groups. They highlight the existence of the different generations, namely the ‘monopoly’, the ‘television; and the ‘highlight generation. Within those three generations fan connections are of main importance and needed for sustaining a brand. Wengler, Chatrath & Berlin (2009) apply market segmentation to the German professional füßball Bundesliga. They argue the relevance of segmenting sport consumers on benefits, and perform a limit-conjoint analysis. They are the only researchers, known by the author, that apply this form of segmentation to the spectator’s market. The respondents evaluated nine product offerings on four attributes: the price, the seat location, the opponent and match attendance. The number of segments identified is four; ‘the aficionados, the show & entertainment addicts, the seat potatoes and the price-sensitives’.

(14)

TABLE 2: Literature Overview on Fan Segmentation

Authors Findings article Focus criteria Segmentation bases Method [sample size: N] Mahony, Madrigal &

Howard (2000) Truly loyal, Spurious loyal, Latent loyal & Low loyal fans Psychographic (attitudinal/behavioral) Direct/prod.specific + Derived/general Factor analysis Survey [N=385] Samra & Wos (2014) Temporary, Devoted &

Fanatical fans Psychographic (motivations) Derived/general Conceptual + Survey [N=460] Smith & Stewart (1999) Passionate partisans, Champ

followers, Reclusive partisans, Theatregoers & Aficionados

Psychographic

(perceptions) Derived/general Conceptual Literature study

Wann & Branscombe (1993), Mullin (1993), Kahle (1996) and Sutton et al. (1997)

Type 1 fans

[internalized, focused, vested], Type 2 fans

[self-expressive, committed] Type 3 fans

[care-free casual, social]

Psychographic

(human values) Derived/general Conceptual Literature study

Rein et al. (2008) Indifferent fans, Eyeballs, Collectors, Attachers, Insiders, The Ensnared

Psychographics within a

Demographic group Direct/general + Derived/general Focus groups, Panel, Projective + Survey

Wengler, Chatrath & Berlin

(2009)* Aficionados, Show addicts, Seat potatoes & Price sensitives

Demographic +

Psychographic (interests) Derived/prod.specific + Direct/general Benefit segmentation. Survey [N=781] Luna-Arocas & Tang

(2007) Leisure-oriented exercisers, Social entertainers, Affective users, Enthusiasts & Passives

Demographic +

Psychographic (interests) Direct/general + Derived/general Cluster analysis Survey [N=218] Tapp & Clowes (2002)* Casuals, Regulars & Fanatics Demographic +

Psychographic (behavior

& lifestyle)

Derived/prod.specific +

Direct/general Factor & Chi-square analysis. Survey [N=667] + Interviews [N=25]

Ferrand et al. (2014)* Show-business lovers, Passionate fans, Admirers of celebrities and fair play, & Event followers

Demographic + Psychographic (brand

association)

Derived/prod.specific +

Direct/general ANCOM-2) Interviews [N=30]

Ferrand & Pages (1996), Quick (2000), Boyle & Haynes (2000), Nash (2000)

Type 1 [irrational, traditional, core, old] &

Type 2 [rational, modern, corporate, new]

Demographic +

Psychographic Derived/general + Direct/general Conceptual Literature study

Hunt et al. (1999) Temporary, Local, Fanatical,

Devoted & Dysfunctional Demographic + Psychographic Derived/prod.specific + Direct/prod.specific Conceptual Literature study Prayag & Grivel (2014)

This study (2017) Indifferent, Enthusiast, Socializer, Competitive* Fanatic-citizens, Price-sensitives, Committed-critics, Seat-potatoes, Social-fans, Ambience-seekers, Social-fans, Joe-public, & the Outsiders Demographic + Psychographic Demographic + Psychographic (Commitment) Derived/general + Direct/general Derived/general + Derived/prod.specific + Direct/general

Hierarchical cluster analysis Survey [N=310]

Hierarchical cluster analysis Benefit segmentation Survey [N=852; 1003]

(15)

Tapp & Clowes (2000) examine the football fans of a single club. By combining qualitative and quantitative methodologies they investigate the complex set of relationships associated with football fans. The qualitative phase allowed the authors to identify the key issues for football fans, while in the quantitative phase statistical procedures were performed. An interesting new finding of their study was that they identified the ‘professional wanderer’ as a subtype of the ‘casual supporter’. This group is more interested in ease of access, payment methods, and entertainment than in the team winning the match. This group was not identified before in previous studies.

Ferrand et al. (2014) describe the UEFA Champions League fans’ as a function of their cognitive content and structure. Four segments were identified (show-business lovers, passionate fans, admirers of celebrities and fair play, and event followers). These groups of fans associate differently with brands. They provide relevant information to efficiently develop marketing strategies that are specifically designed for them. Furthermore, they suggest that it could be profitable to create a framework for strategically managing brand positioning based on brand association networks. The four segments differ in behavior and attitude concerning media, the show aspect, sports stars, top clubs, and performance.

Hunt et al. (1999) argue that fans should be classified, as marketers can leverage this information to develop the best product, distribution, and content mixes. The specific types can be reached more efficiently, because of their varying motivations and behavior. They argue that fan typologies are a fundamental basis for segmentation strategy. For example, to reach a temporary fan, timing is key. This type might be persuaded to attend a match, or buy a pay-for-view package when convinced of the uniqueness of the event (“once in a lifetime opportunity”). Whereas the devoted fan should be approached with information about personalities, and team, league or general sport information. Whereas, Prayag & Grivel (2014) highlight that not a single motive, but rather a combination of motives, drives the choice of going to an event. Understanding that motivation is essential in implementing effective targeting strategies and market segmentation. These motives also inform successful advertising campaigns and promotions.

(16)

pre-Professional Football Fan Typologies: A Latent Class Analysis.

16

specified and in determining an optimal number, multiple distance measures can be used. Disadvantages are the complex and confusing interpretation of the hierarchy. Also, after the data points are grouped into a node, the deterministic nature of this technique prevents reevaluation of this grouping. It is therefore important that to effectively segment, focus should not lie solely on these empirical methods. Wengler, Chatrath & Berlin (2009) proposed a benefit segmentation, but this article isn’t publicly published, and apart from the results not much of their study is known. Furthermore, the input of these empirical analyses is mainly survey data, with samples between 300 and 780 respondents. The conceptual studies are based on literature research and are useful for providing a general and combined overview of multiple frameworks. A disadvantage of these studies however is the lack of adding new insights based on empirical research. Summarized, most authors study a combination of general psychographic and demographic variables as the bases for segmentation, either conceptually or empirically with a traditional method.

2.4. Research Model

Ultimately, based on this previous research and corresponding evidence from the academic literature, it can be concluded that for the case of segmenting supporters it is interesting to use both direct, derived, general and product-specific segmentation criteria. Derived bases to segment (active) on and direct observable bases to describe (passive) different homogeneous groups. General segmentation bases are good in bringing a segment to life, whereas product-specific bases are good at explaining actual behavior. Furthermore, many organizations are interested in maximizing shareholder return and like to know the actual value that consumers are willing to pay in return for the sacrifices that they are willing to make. In this study, I therefore propose to segment on both the consumer’s psychographics as on the benefits sought by consumers. As lifestyle values are interesting to know, it is in this case also interesting to identify the actual consumer requirements (wants and needs). The previous studies’ findings, criteria, bases and methods together, indicate the interest and relevancy of measuring preferences and see whether the compared findings show overlap.

(17)

17

FIGURE 1: Research Model

(18)

18 3. Methodology & Empirical Framework

First, a brief overview of the company FC Groningen B.V. is provided. Next, the research design of study (1) and study (2) are discussed. Study (1) includes a hierarchical cluster analysis to measure supporters’ psychologic and demographic values, while study (2) consists of a choice-based conjoint analysis used to measure supporters’ preferences by measuring their utilities. These utilities are assumed to be maximized and lead to a rational decision by each respondent.

3.1. FC Groningen

The proposition of a heterogeneous fan base is tested against data from supporters from the Northern-Netherlands professional football club FC Groningen. The club is founded in 1971, plays in the Dutch Eredivisie league and plays in the ‘Noordlease’ stadium. This stadium has a maximum capacity of 22.550 people (with a weekly occupancy rate of ±75%). Its colors are green-white and they present themselves as ‘the proud of the North’. Whether good or bad times, it aims to create an environment where players and supporters perform together. With its national image and European ambitions, it aims to create a commitment to society in the field of region promotion, social responsibility, health, development and the economy. All of this is based on a fundament of five core club values: Unconditional, ambitious, committed, accessible and combatable. Thus far, the club gathers (demographic and revenue per customer) data from its fans when those are creating online accounts or buy tickets/season cards online. For next seasons, they plan to collect data by introducing a new card-based paying system, and electrical gateways at the stadium entrances.

3.2. Study (1)

3.2.1 Design Study (1): Hierarchical Clustering Survey

(19)

19

3.2.2. Segmentation Method

For this cluster analysis, I measure the (dis)similarity of customers based on their answers to the survey. Supporters with similar characteristics are group together. According to numeric and strategic criteria a number of segments is selected, whose needs are profiled using the cluster means.

To tackle the first problem of deciding on which variables to group the respondents, the variables are formulated as either active (used for clustering) or passive (used for group identification). The variables are perceived characteristics from the products or the customers. To prevent problems with scale differences, all variables are standardized/transformed to have a mean of 0 and a variance of 1 (Z score). The active variables and segmentation bases in this study are “Commitment” (psychographic dimension) and “Age” (demographic dimension). The commitment variable is a combination of questions concerning the ‘Role’ of FC Groningen, the frequency of following ‘News’ about FC Groningen, and the amount of ‘Conversations’ about FC Groningen. All were converted to a Z-score and subsequently composited to one factor measuring commitment. This procedure is recommended by Ackerman & Goff (1992), who argue that when the goal is to provide stable measures of the underlying abilities, a composite should be formed with (unit-weighted) Z-scores of constituent tests.

(20)

20

3.2.3. Model Specification Study (1)

For the determination of segments in study (1), I estimate a hierarchical cluster analysis using Ward’s linkage method with a Euclidean distance, as mentioned in the previous section.

To measure the quality of clustering, the sum of the squared error (SSE) is used. In other words, the error of each data point is calculated (its Euclidean distance to the closest centroid). These are used to compute the total sum of the squared errors. The clusters with the smallest squared errors are preferred, since this means a better representation of the centroids of those clusters. The SSE is formally defined as:

𝑆𝑆𝐸 = ∑𝐾_𝑘=1∑_𝑥∈𝐶_𝑘𝑑𝑖𝑠𝑡(𝑐_𝑘, 𝑥)2 (1)

Where

dist

is the standard Euclidean distance between two objects in the Euclidean space and k are the different segments (𝑘1, 𝑘2…K). It finds the distance between the cluster mean

c

and

explanatory variable x. Ultimately, the centroid minimizing the SSE of each cluster is the mean. This centroid (mean)

c

of the 𝑘𝑡ℎ cluster is defined formally as:

𝑐_𝑘 = 1

𝑚𝑘

∑_𝑥∈𝐶_𝑘𝑥 (2)

3.2.4. Sample Characteristics Dataset (1)

To get an initial feeling with the data, I will provide a summary and some visualizations. In the dataset for study (1) “Age” and “Commitment” (an overlapping variable for Role, News and Conversation) are the active variables in determining the number of segments. I describe this dataset based on their numeric mean, median, standard deviation, minimum and maximum values.

Dataset (1) consists of a total sample of 852 respondents, having observations of 13 variables. The data were collected between February 1st_{2017 and march 1}st_{2017 via the supports’ club accounts.} For each respondent, I have information regarding:

• Psychographics: Role; Conversation; News; Drivers • Demographics: Gender; Age

(21)

21

The questions are distributed via the club’s website when logged in with a my-club account. All categorical variables with multiple levels are recoded as new dummy variables, each having a value of [0:1]. Now, they are recoded into two categories indicating either the absence or the presence of a certain characteristic. Furthermore, to make a comparison between the segments, all variables are set to a ‘common scale’. They are transformed into Z-variables varying from zero to one. This way, relative interpretation will be possible, and certain characteristics won’t be more important than others because of their absolute scale.

The data can be identified as cross-section data. To get a better feeling with the dataset, a summary of the descriptive statistics is provided in table 3. As can be observed, there are some patterns in the data. 84% of the respondents are male (vs. 16% females). The average age is 41.6 years. On average, 13% of them are in possession of a club card, 53% of a season card, 99% of a my-club account and 23% buys loose tickets in the open sale. The amount of variation in the commitment indicators is ± 2. In the media variable it this is 8.81, indicating that most respondents (64%), watch between 1-19 matches online or via television each season.

TABLE 3: Descriptive Statistics Dataset (1)

[N=852] Mean [𝝁] Median S.d. [𝝈] Min. Max.

Role 6.60 7 2.04 1 10 News 7.81 8 1.80 1 10 Conversation 7.24 8 1.76 1 10 Media 10.21 10 8.81 0 34 Age 41.6 43 15.85 8 80 Gender 0.84 - - 0 1 Club card 0.13 - - 0 1 Season card 0.53 - - 0 1 My-club account 0.99 - - 0 1 Open sale 0.23 - - 0 1 3.3. Study (2)

3.3.1. Design Study (2): Benefit Component Survey

(22)

22

presented on which respondents are asked for their most preferred option. The choice selections are used as the dependent variable and interpreted as a natural manifestation of preference. Furthermore, respondents can indicate not to choose any, with the inclusion of a “None” option. This survey was distributed via the season cardholder newsletter. They had to assume the for them realistic situation, where they faced the choice of extending their season card for next season. People were faced with ten choice sets, each with different combinations of attribute levels (see appendix 7.1). Each time, they had to choose the one they preferred the most to gain insights in relative preferences. Furthermore, to gain more general insights, respondents were asked some demo- and psychographic question at the end of the survey. To motivate people to fill in the survey, a price (meeting with the selection players after the match) was promised to a random respondent. The attributes and their levels on which people had to base their choices are discussed below. In total, 1003 fans filled in the survey.

3.3.2. De-compositional Choice-Based Conjoint Analysis

The season offerings are perceived as attribute bundles (5 factors), where the utility of a season card equals the sum of the utilities of the individual attribute levels (features). These preferences for certain attribute levels are statistically decomposed from the overall alternative evaluation. The selected best choices are used as the dependent variable. In the experimental design, the offerings are broken down into independent attributes and levels. These attribute level combinations are shown to the respondents. An overview of these are presented in table 4. The attributes are assumed to be independent (levels can combine freely). The levels have a concrete meaning and are assumed to be mutually exclusive from the others. The argumentation of using these specific attributes is based on previous literature. A summarizing overview can be found in appendix 7.2.

TABLE 4: Attributes/Levels Overview CBC Analysis

Attributes Levels

Atmosphere (in the stadium) Sold out (100%), Regular occupation (80-90%), Empty seats (70-80%) Facilities (menu + waiting time) Excellent [<1 min.], Regular [1-5 min.], Limited [5-10 min.]

Seat location Long side, Short side, 2nd_{ring: Short side, 2}nd_{ring: Long side}

Security & Enforcement Good [1-5 incidents], Average [5-10 incidents], Poor [>10 incidents] Team performance (rank) Very good [1-4], Good [5-8], Average [9-12], Poor [13-18]

(23)

23

In the factorial design, a subset of the full factorial design is used. For the design to be efficient, the subset is balanced (each level is displayed an equal number of times) and orthogonal (each level combination appears an equal number of times and there is no correlation between attributes). In the choice design, an allocation of the stimuli to show in the different choice sets is made. This allocation shows minimal overlap and non-dominated choice sets. In the implementation phase the stimuli and choice sets are presented to respondents in the survey. For visualization purposes, an image is added for the atmosphere attribute. This is based on the findings of (Eggers, Hauser, & Selove, 2016), who argue that using craft, e.g. high-quality images and incentive alignments, lead to increased predictive validity. The survey includes an incentive alignment where a random respondent wins a meeting with the selection of FC Groningen after the first match of next season. Furthermore, a no-choice option is added if respondents are not interested in the offering.

3.3.3. Model Specification Study (2)

For the determination of segments in study (2), I estimate a Latent Class (LC) model with K-segments. This kind of analysis is suggested as an effective model-based tool for market segmentation (Wedel & Kamakura, 2000). Latent (unobserved) class models overcome the limitations of a-priori segment selection and that of higher-level aggregation, since they allow for unobserved heterogeneity between individuals., by treating the underlying segments as hidden or latent classes (Vermunt, 2003). Models for different segment solutions are estimated and the model that fits the data best will be selected. In these estimations, the model parameters per segment and the number of segments are calculated simultaneously. When this approach is applied to the gathered survey data, I must specify the, unknown, actual number of segments K. The result will be a probabilistic classification of consumers into different classes. The model uses a log-likelihood function, which is maximized with the expectation maximization (EM) algorithm. This algorithm finds the set of β’s that maximizes the likelihood function (Vermunt, 2003). All supporters attach a part-worth utility to each attribute. This model assumes that the ideal offering is a combination of the attributes included in the survey, and that people maximize their utility. The regression parameters are estimated as a random utility model of individual i for the different alternatives j, consisting of a systemic (V) and a stochastic (𝜀) utility component, following a logistic distribution:

𝑈𝑖𝑗 = 𝑉𝑖𝑗+ 𝜀𝑖𝑗 (3)

(24)

24

In words, I estimate the probability of observing a response pattern s, which is a combination of segment k’s response on the latent variable X and the probabilities of observing a combination of responses 𝑠𝑘 conditional on the respondent’s classification. Econometrically, this looks like:

𝑃(𝑌𝑖𝑗 = 𝑠) = ∑𝐾𝑘=1𝑃(𝑋𝑖𝑗 = 𝑘)𝑃(𝑌𝑖𝑗 = 𝑠|𝑋𝑖𝑗 = 𝑘) (4)

In this probability (P) model,

𝑌_𝑖𝑗 = the response of individual i, within alternative j, where 𝑌𝑖 = {

0 𝑖𝑓 𝑈_𝑖 ≤ 0 1 𝑖𝑓 𝑈_𝑖 > 0; 𝑋𝑖𝑗 = the explanatory latent class variable; including a vector of regression parameters in segment

𝑘 = a particular latent class, with a total of K latent classes;

𝑃(𝑌_𝑖𝑗 = 𝑠) = the probability of observing a response pattern s; which is a weighted average of 𝑃(𝑌_𝑖𝑗 = 𝑠|𝑋_𝑖𝑗 = 𝑘) = the class-specific probabilities.

Furthermore, the dependent variable is a discrete/categorical choice selection dummy, with multiple choices. I assume that the supporters choose the alternative with the highest utility. E.g. when choosing between two choice sets, the first alternative will be selected if 𝑈1 > 𝑈2. For each

attribute, the models are specified as part-worth. Nominal attributes (Atmosphere, Catering, Safety, Team performance) are always part-worth. For the price attribute, I perform a Chi-squared test to formally test whether the model should be specified as fully part worth or as one with a numeric price variable. This test is significant (Chisq=221.02, df=1, critical value=3.84, p<0.001), indicating that a price linear vector and a part worth vector model are significantly different, in favor of the more complex part worth model. Therefore, the price attribute is also specified as part-worth. Summarized, the utility function U is a linear combination of systematic part-worth utilities V with a random error term. The dependent variable (the alternative selection dummy) can exhibit multiple states. The chosen options can be any alternative from choice set J.

3.3.4. Classification

1 ,

k

k i i

(25)

25

3.3.5. Validation: Assessment and Comparison of Fit

Essential aspects of proper model selection are validation measures regarding model assessment and comparison. I will examine each models’ goodness of fit. An important measure in this is the log-likelihood ratio test. Since this instrument on its own keeps improving when adding more segments, it is proposed to measure multiple Information Criteria (IC) used to compare different models with a different number of latent classes. These IC measure the models’ complexity (mostly based on the log-likelihood), by assigning different penalty terms to each of them. The criteria seek to incorporate the divergent considerations of accuracy of estimation and the ‘best’ approximation to reality. I compare the estimated models based on two different IC measures. These indicate the best data fit when they are minimized. The principle of parsimony suggests a trade-off between variance and bias as the number of free parameters in a model increase (Anderson, Burnham, & White, 1998). For this latent class analysis, the Bayesian Information Criteria (BIC), and the Consistent Akaike Information Criteria (CAIC), are preferred, since those criteria better fit large sample sizes. Also, they penalize more for a greater number of parameters and latent classes than e.g. the AIC and AIC3. The BIC penalize complex models, depending on the sample size, where more parsimonious models are favored. The CAIC are often selected when prediction is the goal of modelling. It assumes that the order of the (true) model does not change when the sample size increase. These criteria are preferred for this large sample size, since they reduce the tendency of overestimation by other criteria. In addition, the models are also compared and evaluated based on their adj. Pseudo R^2, and measurement errors. Finally, the Independence of Irrelevant attributes is an important assumption which is many times violated in these types of models, however the LC specification resolves this problem. For further details on this issue, see Vermunt (2003).

3.3.6. Sample Characteristics Dataset (2)

In the second dataset used to test possible predictors of fan heterogeneity, a “Selection_dummy” of the different choices with value [0;1] is the dependent variable. I describe the data based on their numeric mean utilities.

(26)

26

• Benefits: Atmosphere, Catering, Seat location, Safety, Performance & Pricing preferences • Demographics: Gender; Age; Education; Income; Marital status; Kids

• Behavioral: Commitment; Knowledge; Frequency; Extending preferences • Company and Transport

On average, 83% of the respondents are male (vs. 17% female), with an age of 41.8 years old. Relatively, most respondents are well educated (59.6% finished HBO or university), have a high income (29% makes >3500 net p.m.), are married (54%), and have 1-2 kids (53%). Furthermore, they normally decide to extend their season card as soon as it is possible (43%). Overall, almost all respondents are (very) committed to the club and have moderate to much knowledge about soccer. Finally, they attend almost all matches, mostly by car (57%) and with their friends (40%). An overview of the mean values per attribute level is provided below in table 5. A decreasing pattern in level preference can be observed. Furthermore, ‘Atmosphere’, ‘Performance’, and ‘Price’, show the greatest range which indicate to be the most important predictors, whereas ‘Catering’ shows the smallest. For the price attribute, on average, respondents seem indifferent between the first two levels, while paying €370 is to a high extent not preferred, compared to the other levels.

TABLE 5: Descriptive Statistics Dataset (2) Attribute [N=1003] Mean Mean Atmosphere Safety Occupancy of 100% 0.53 1-5 incidents 0.19 Occupancy of 80-90% 0.11 5-10 incidents 0.01 Occupancy of 70-80% -0.64 >10 incidents -0.19 Catering Performance Extended 0.04 Ranked 1-4 0.52 Regular 0.04 Ranked 5-8 0.22 Limited -0.08 Ranked 9-12 -0.17 Ranked 13-18 -0.58

Seat Location Price

Short side 0.17 €150 0.36

Long side 0.03 €220 0.35

Short side: 2nd_ring _{-0.19 €270} _-0.03

Long side: 2nd_ring _{-0.01 €370} _-0.69

None_option -0.93

(27)

27 4. Estimation Results Empirical Analysis

4.1. Discussing the Results

First, the results of both clustering methods will be presented and discussed. Both data sets show an optimal number of six segments. Second, I will examine if both methods show some overlaps. Finally, I present a robust fan segmentation framework based on the results.

4.2. Study (1): Hierarchical Clustering Method

The clustering method performed on dataset (1) indicates that there are six segments describing the data best. The segments are formed based on the active variables “Commitment” and “Age”. The passive variables, used for describing the groups, are “Card_info”, “Transport”, “Company”, “Drivers_to_go_to_the_stadium” and “Gender”. The results are visually represented below in figure 2.

FIGURE 2: Clustering on the Age and Commitment Dimensions

Note: From left to right the segment sizes are: 20.5; 7.5; 13.6; 14.1; 26.9; & 17.5% respectively

From the figure, it can be observed that there are three segments showing a high(er) commitment, and three segments showing a low(er) commitment. These are all spread out over a demographic age dimension. The bubble size indicates the relative segment size. Interestingly, while there are segments representing most age groups, there is no segment describing the group 30-45 well. This might be explained by the fact that this group has a busy lifestyle, compared with the other age groups, and has less time to fill out the survey. Furthermore, it does not mean that there are no target-consumers in this category, but they might be committed to a lower extent. When someone is decreasingly committed, they might decide not to fill in the questionnaire, and there is no/little information about them. This doesn’t make them not interesting, but a potential target group. It is interesting to see if this age group is also absent in the results of study (2). The main results are reported in table 6, whereas the detailed segment results can be found in appendix 7.4.

(28)

28

TABLE 6: Main Results Clustering “Hierarchical Euclidean Distance Method”.

(29)

29

4.3. Study (2): Benefit Segmentation Method: Conjoint Analysis

To observe homogeneous response patterns for the different attributes and to study possible similarities or differences within the second sample of 1003 respondents, I applied the multinomial logit latent class analysis as described in the model formulation section mentioned previously. For a different number of supporter segments, alternative parameters are estimated. Before interpreting the estimation results, I validate the different options and propose the best model based on those. Since the optimal number of segments is not known prior to the analysis or retrieved by the estimation method, I estimated models for a several number of segments (K=1…15). Next, I find the best fitting model according to log-likelihood-based measures (Information Criteria), the number of parameters, adjusted pseudo R2’s, classification errors and the ease of workability and interpretability. A total of 15 different models were run, each with a different class solution. An assessment and comparison of model fit between the different models is reported in table 7.

TABLE 7: Validation Criteria Different Models

Model LL BIC CAIC Npar Df Class.err Pseudo 𝐑𝟐

1-class -13.481,8 27.073,7 27.089,7 16 959 0.0000 0.23 2-class -12.358,9 24.944,8 24.977,8 33 942 0.0361* 0.29 3-class -12.057,8 24.459,7 24.509,7 50 925 0.0857* 0.31 4-class -11.800,9 24.062,9 24.129,9 67 908 0.1024 0.32 5-class -11.575,9 23.729,9 23.813,9 84 891 0.1098 0.34 6-class -11.368,1 23.431,4 23.532,4 101 874 0.0997* 0.35 7-class -11.215,9 23.244,0 23.362,0 118 857 0.1055 0.36 8-class -11.084,5 23.098,1 23.233,1 135 840 0.1096 0.37 9-class -10.960,1 22.966,4 23.118,4 152 823 0.1084 0.37 10-class -10.871,9 22.906,9 23.075,9* 169 806 0.1168 0.38 11-class -10805.4 22.890,9 23.076,9 186 789 0.1238 0.38

Note: * represents the optimal number of segments based on that specific column’s measure. The BIC value gets to an absolute minimum at a 14-class solution.

Based on the validation measures, we see that the data fits better to a multiple-class model than to an aggregate model (1 homogeneous class). The log-likelihood ratio decreases as the number of segments increase, indicating that the model improves compared to a null model and better fits the data, as more segments are added. This supports the proposition of a heterogeneous fan base.

Concerning the Information Criteria, also the BIC, and the CAIC decrease towards the minimum, as the number of segments increase. When examining these scores from the table, I observe a minimum value at 10-classes regarding the CAIC, and at 14-classes regarding the optimal BIC. However, the marginal improvement of the criteria flattens of at 6-classes. This is visually represented in figure 3. Relatively, the models do not improve a lot going from a six- to a ten-class solution, and not many extra information is extracted when adding these extra segments.

(30)

30

Combined with a measurement error below 10% (9.97%), a pseudo adjusted R^2 value of 35%, better interpretability, and the fact that a lower-class model is more parsimonious (less complexity), I conclude that the “optimum” number of segments representing the data best in this case is six. Furthermore, I test different models for the six-class solution to use a most optimal model representing the supporters. First, I estimate a model with price as a linear attribute instead of part-worth. Furthermore, I estimate a model including covariates in determining the latent classes. Inclusion of covariates can help to better understand segments in terms of demographics and other external variables. Also, they help classifying new cases into the appropriate segments (Vermunt, 2003). The relevant statistics are provided below and indicate that a part worth non-covariate vector model is best. The inclusion of the significant covariates (education, age, extending behavior and commitment), resulted in the best model fit concerning the different combinations of covariates. Although significant, the model fit is still worse than a variant without any covariates, and so these characteristics will be used as passive variables in describing the segments rather than active segmentation criteria. The different model comparisons are reported below, in table 8.

TABLE 8: Comparison of Several Six-class Models

Models LL BIC CAIC Class.error Pseudo R^2

(1) 6_class_Partworth -11.368,1 23.431,4 23.532,4 9.97% 0.35

(2) 6-class_PriceLin -11.552,6 23.717,8 23.806,8 10.64% 0.34 (3) 6_class_Covariates_All -11.263,0 24.962,4 25.316,4 8.88% 0.36 (4) 6_class_Covariate_Edu+Age+Extending+Commitment -11.408,9 23.946,6 24.110,6 9.02% 0.35 (5) 6_class_Covariate_Age -11.501,4 23.649,7 23.743,7 10.60% 0.34

(31)

TABLE 9: Parameter Estimates of the Choice-based Conjoint analysis.

Note:・p<0.10, *p<.0.05, **p<0.01, ***p<0.001.

Category Attributes Class[1] Class[2] Class[3] Class[4] Class[5] Class[6] Wald p-value Wald(=) p-value

100% 0,6952 0,4862 0,7022 0,2897 0,4441 27,835 12,422,40*** 1,4e-258 1,999,76*** 1,6e-37 Atmosphere _80-90% _0,1919 _0,3741 _0,2004 _0,0092 _0,1621 _-0,6875 70-80% -0,8871 -0,8603 -0,9026 -0,2989 -0,6063 -20,960 Extended -0,0367 0,1217 0,0691 0,0438 0,0918 0,3201 570,59*** 7,7e-8 184,38** 0,038 Catering _Normal _0,0777 _0,0507 _0,0843 _0,0118 _0,1007 _-0,1632 Limited -0,0410 -0,1725 -0,1534 -0,0556 -0,1925 -0,1569

Long side -0,1186 -0,1145 0,2840 0,9734 0,4827 -0,1312 5,183,54*** 1,4e-98 4,053,58*** 5,2e-77 Short side 0,3240 0,2427 -0,1424 -0,4626 0,1558 0,0757

Seat Location ₂_nd_{ring: long} _-0,0193 _-0,0056 _-0,2490 _-0,9112 _-0,3948 _0,2205 2nd_{ring: short} _-0,1861 _-0,1226 _0,1074 _0,4004 _-0,2437 _-0,1650

(32)

4.3.1. Relative Importance

The parameters showed in table 9 cannot be compared directly, but should be compared on a common scale. Therefore, the relative attribute importance is provided below in table 10. Some great differences between segments can be observed. For example, class 1 considers price and atmosphere the most when considering extending their season card, while almost not basing their choice on catering and seat location. Whereas, e.g. class 4 bases their choice to extend mostly (45%) on the location of their seat, and finding team performance of less importance (12%). Class 3 indicated the importance of sportive success (52%), while class 6 almost only cares about a good atmosphere (55%). Furthermore, class 6 doesn’t seem to care about safety in their decision at all.

TABLE 10: Relative attribute importance, grouped by class

4.3.2. Profiles study (2)

As with the hierarchical clustering method, the segments are described with abovementioned active, and certain passive (Age; Education; Commitment; Extending behavior; Gender; Income; Transport; & Company) variables. These findings are reported below in table 11. As with the clustering method, some great similarities and differences in describing the different segments can be observed. Based on these demographic, psychographic and behavioral segmentation bases, each group is given a name. The six (committed) fan profiles are the ‘Fanatic-citizen’, ‘Price-sensitive’, ‘Committed-critic’, ‘Seat-potato’, ‘Social-supporter’ and the ‘Ambience-seeker’. Next, I will compare both methods and observe the similarities/differences across them. All show a high commitment, which was expected since the study is only performed on fans with a seasonal card. Segment 1: the “fanatic-citizen” is ‘diverse’ concerning preference and demographic variables. Relatively, of all groups they are represented the most by women (22%), and the respondents are divided over various age groups. They show a high commitment and multiple attributes play a role in buying a season card. Contradictory to e.g. the static segment 4: the “Seat-potatoes”, consisting of mainly older males, coming to the stadium alone or with their partner for a good seat. Segment 1: the “Ambience-seekers” are relatively young and seek atmosphere in the stadium when they attend a match day.

Class [size]: 1 [24%] 2 [24%] 3 [18%] 4 [16%] 5 [9%] 6 [9%] Aggregated

(33)

33

TABLE 11: Parameter Estimates Choice-based Conjoint Analysis.

[24%] [24%] [18%] [16%] [9%] [9%] N=1003 “Fanatic-citizen” “Price-sensitive” “Committed-critic” “Seat-potato” “Committed-social” “Ambience-seeker”

Drivers* Demographics Price (++) Atmosphere (±) Team performance (±) Price (++) Atmosphere (±)

Performance (++) Seat Location (++)

Safety (±) Atmosphere (+) Seat Location (±) Price (±) Atmosphere (++) Age* 15-35 (64%) 25-55 (61%) 25-55 (63%) >55 (48%) 40-55 (45%) 15-35 (59%) Education* Medium (45% hbo; 27% mbo) Lower (36% hbo; 30% mbo) Highest (49% hbo; 24% uni) High (41% hbo; 23% uni) High (38% hbo; 26% uni) Medium (40% hbo; 36% mbo) Gender 78% male 81% male 86% male 86% male 83% male 89% male

Income

Psychographics

Lowest Low Highest Average Average Low

Commitment*

Behavioral

High (7.5) High (7.6) High (8.1) High (6.8) High (7.4) High (7.9)

Extending* End of the comp (35%) A.S.A.P. (35%)

A.S.A.P. (49%) A.S.A.P. (51%) A.S.A.P. (45%) When friends (20%)

Selection known (21%) End of the comp. (29%)

When friends do (24%) A.S.A.P. (40%) Transport Car (52%); Bike (38%) Car (60%); Bike (32%) Car (61%); Bike (32%) Car (62%); Other (38%) Car (64%); Bike (25%) Bike (32%); Bus (11%) Company Others (20%); Kids

(24%); Friends (41%) Kids (20%); Friends (42%) Kids (22%); Others (19%) Partner (10%); Alone (16%) Kids (31%); Friends (36%); Partner (13%) Friends (49%); Other (25%)

Note: *The segments are significantly different between the Drivers, Age, Education, Commitment and extending behavior.

4.4. Segmentation Framework Study (1) & (2)

When comparing the results of both methods, I observe some similarities and some differences between the segments found in the (hierarchical clustering) and in the [benefit segmentation]. I do this by reporting a function matrix in which both methods are integrated in one overarching framework. This framework is presented in the following section.

Main Differences:

(34)

34

Main Similarities:

✓ Both segment (4) and [4] are old (55+), committed, in possession of a seasonal card, come with the car, in company of their partner or alone. They both value facilities, safety and live players the most. This is translated into a good seat location in the stadium. They are identified as the “Seat-potatoes”.

✓ Both segment (5) and [5] are older (40-55), committed, in possession of a seasonal card, come with the car, in company of kids or alone. They both value atmosphere the most, and the other characteristics are in an equal range. This latter corresponds with the social aspect (social interactions are more important than a specific characteristic). They come with their partner, friends or kids. Therefore, they are identified as the “Committed-socials”.

✓ Both segment (6) and [6] are young (15-35), committed, in possession of a seasonal card, come with the bike, in company of friends and others. They both value the atmosphere the most as a driver to come to the stadium. They are therefore identified as the “Ambience-seekers”.

(35)

35

TABLE 12: Supporter Segmentation Framework Hierarchical >

Benefit Outsider Social fan Joe-public Seat-potato Committed-social Ambience-seeker Fanatic-citizen

Price-sensitive Both Low & High Committed Fans Committed-critic

Seat-potato

Committed-social The High Committed Fan

Ambience-seeker

The segments are combined in an overarching framework indicating the similarities and differences discussed previously. Ultimately, the identified segments are combined and described as:

Segment [1]: the “Outsider”. The older, low-committed fan. They come to the stadium, with varying company, to see an appealing and comfortable match.

Segment [2]: the “Social-supporter”. The younger low-committed social fan. They are appealed by the price of an open-sale ticket, the atmosphere and prefer a good catering service most compared to other segments. They mostly come with friends and family.

Segment [3]: the “Joe-public”. The average-aged neutral fan. They come to the stadium for social interactions, mainly with their kids, friends or business partners.

Segment [4]: the “Fanatic-citizen”. The younger-than-average-committed fan that prefers a mix of price, atmosphere and team performance. They come to the stadium with their family and friends.

Segment [5]: the “Price-sensitive”. The average-aged-committed fan, that buys a season card when it is attractively priced. They prefer the stadium’s atmosphere and attend matches with kids or friends.

Segment [6]: the “Committed-critic”. The average-aged-committed fan that wants success, and does not care about paying a higher price for it. They are identified as loyal and are in the highest income group.

(36)

36

Segment [8]: the “Committed-social”. The older-committed fan that comes to the stadium with friends, their partner, alone or with kids and don’t value a specific match-attribute much higher than the rest.

Segment [9]: the “Ambience-seeker”. The younger-committed fan that comes to the stadium for sensation and entertainment. They come with friends or other family-members.

4.5. Managerial Implications / Strategy

For management, it is important to target each segment with an efficient and effective strategy. Below in table 13, I will propose a communication target strategy for internal and external policies.

TABLE 13: Strategical Target Approach per Segment.

Tone of Voice Visuals Information Positioning

Outsiders ‘General’ Team photo Arrangement/flyers Top sport & athletes Social-supporters ‘Solid’ Atmospheric moments Social media/adverts Unique arrangements with

friends

Joe-public ‘Solid’ Match moments Flyers/adverts The alternative to substitutes Fanatic-citizens ‘To the point’ Match moments,

related aspects

Price-focused campaign FCG as ‘proud’ & core values

Price-sensitives ‘Low-threshold’ Combativeness Price-focused flyers Open & transparent

Committed-critics Solid Success moments Social media/mail Open, ambitious & entertaining Seat-potatoes ‘Soft’ Enjoyable moments Paper/letter Reliability & comfortability

Committed-socials ‘General’ Amusement Mail/adverts FCG = ‘nice day-out’

Ambience-seekers ‘Low-threshold’ Atmospheric moments Social media/mail Match = an atmospheric event