FAIR TRADE CONSUMPTION

(1)

WHAT ARE ITS DRIVERS?

Master Thesis Leonie Beekman

(2)

Master Thesis

Master of Business Administration: Marketing Management and Marketing Research Faculty of Economics and Business, University of Groningen

Under supervision of Prof. dr. Jaap Wieringa and Alec Minnema Msc.

By Leonie Beekman Tonny van Leeuwenlaan 31 9731 KH Groningen

+31 (6) 12162742 student number: 1529056

Summary: The study at hand is the first to find that the majority of consumers is willing to trade off their preferred level of particular functionality product attributes (accepting a less preferred level of these attributes) for a fair trade label. Apparently, consumers are willing to put aside particular individual benefits for ‘the greater good’; for sustainability in general or for a fair trade label in specific. Prior literature found that consumers are not willing to trade off a good functionality for sustainability and this finding shines another light on the value of the fair trade label. It implicates manufacturers might be able to ignore some functional attributes of their product when producing sustainable products such as fair trade products. Secondly, this study finds a willingness to pay more for the fair trade label, which is in line with prior literature. Furthermore, we are the first to find that an interaction effect exists between the price of a product and the fair trade label. Apparently, a fair trade label has more worth to consumers, has more positive effect on the utility of a product, if it is combined with a low price. This implicates that an optimal price for products with a fair trade label can be calculated.

Third, prior literature is contradicting on the issue of predicting sensitivity to sustainable attributes by observable measures and ethical stances of consumers. The study at hand concludes that socio-demographic characteristics of a consumer are not a driver of the importance he/she attaches to the fair trade label, while ethical concern

is.

(3)

PREFACE

The piece that lies before you is the final product of my studies in Business Administration at the University of Groningen. My graduation process began with an internship at Publicis in Amsterdam, which enabled me to learn a lot about the advertising world. Yet, I also gained insights in my future career interests; I will not be happy working in advertising or in convential Marketing. This lead to difficulties in my thesis and I had a hard time finding motivation to continue with my thesis and the then current subject about international advertising. Therefore, I began orientating on a new subject and came across the Max Havelaar fair trade label. After visiting the office of the Max Havelaar foundation where I met with their communication manager, I had clear I wanted to research the consumption of fair trade products. Yet, I found it challenging to write the theoretical part of this study, as quite a lot seemed interesting and relevant to include and research. However, once data became available to analyse, I really enjoyed working on my thesis. It was exciting and challenging to work with the for me unknown methodology of Hierarchical bayes modelling.

I would like to thank some people for making it possible to graduate. First, a thanks to Publicis and in particular Gertjan Hafkamp, Caroline van Hoof and Mathijs Boonstra for their support and giving me the opportunity to see advertising and marketing in action. Furthermore, I would like to thank my supervisors Jaap Wieringa and Alec Minnema for their support and constructive feedback. Also, my thanks go out to Frank Roffel, who helped me with getting my experiment online. Moreover, a thanks to my friends and family for their help and encouragement. Last, I would like to give a special thanks to my parents, sister and Mathijs who always support me no matter what.

I hope you enjoy reading my thesis!

(4)

1. INTRODUCTION

“Fair trade is popular in Holland. Labels which guarantee farmers a fair price do well. In three years the sales almost doubled”

This was the title of a newspaper article in a Dutch newspaper in October 2011 (Reijn 2011). It refers to a research by GFK, which also found that fifty percent of Dutch households bought a fair trade product occasionally in 2010, while this was only thirty percent in 2008 (NCDO 2011). The growing market is a sign of increasing involvement of consumers towards fair trade issues. Supporting this, consumer behaviour research indicates that consumers are ever more concerned about environmental and social issues and are growingly acting upon these concerns of sustainability in their consumption (Auger et al. 2010, Backus et al. 2011, Distrifood.nl 2011, World Business Council for Sustainable Development 2008, Schuttelaar & Partners 2011). With ethical consumerism, consumers can show their concerns by buying a product that is related to an ethical issue and/or by boycotting products for their negative features (DePelsmacker et al. 2005). As such, there has been increasing activity of consumers regarding the so-called social behaviours of organizations (Auger et al. 2003). An example is the ‘unfair chocolate letters’ campaign for non-sustainable produced chocolate letters by Oxfam Novib in 2009. One hundred thousand Dutch consumers showed their support online and lot of pressure was put on manufacturers and retailers producing and selling these wrong chocolate letters. As a result, 95% of all chocolate letters were sustainable produced in 2010 (Oxfam Novib 2011). Likewise, companies are starting to show more concern, are increasingly undertaking Corporate Sustainability Responsibility (CSR) initiatives and are making sustainability part of their business and strategy (Bhattacharya and Sen 2004, Max Havelaar 2011b, McGoldrick and Freestone 2008, Ministry LNV 2009, United Nations 2010).

1.1 Problem statement

The research question of the study at hand is the following: What are the drivers of fair trade

consumption?

(6)

societal-based labels have received is modest and they are not often studied (McEachern 2008, Moussa and Touzani 2008). Additionally, the importance of these attributes and their influence on consumer behaviour and decision-making is not clear (Tagbata and Sirieix 2008, Moussa and Touzani 2008). Also, in general, consumer decision making towards sustainable food consumption and the willingness of consumers to consider non-product social features in their purchase decision is much discussed but little understood (Auger et al. 2008, Vermeir and Verbeke 2006). Moreover, it is found that the found attitudes and intentions of consumers towards sustainability are not consistent with actual purchase behaviour, known as the ‘attitude-behaviour gap’ (Auger and Devinney 2007, Bird and Hughes 1997, Carrigan and Attalla 2001, Cowe and Williams 2000, DePelsmacker et al. 2005, Nicholls and Lee 2006, Roberts 1996, Vermeir and Verbeke 2006).

1.2 Theoretical, managerial and social relevance

The study at hand examines consumer buying intentions and consumer decision making of sustainable products and contributes to the understanding of sustainable buying behaviour of consumers. Furthermore, this study examines the importance of the fair trade label in consumer decision making and hereby it adds to the currently little knowledge.

Prior literature found that consumers are not willing to trade off a good functionality for sustainability. In the study at hand is the first to find consumers are, to some extent, willing to trade off their preferred level of particular functionality attributes in a product in order to obtain a fair trade label. This shines another light to the value of the sustainable attribute; the sacrifice of functionality consumers are willing to take (Devinney et al. 2010).

Furthermore, this study is the first to find an interaction effect between the price of a product and the fair trade label. This indicates that a fair trade label has more worth to consumers, has positive more effect on the utility of a product, if it is combined with a low price. Moreover, we find a willingness to pay more for the fair trade label, which is in line with prior literature. Additionally, the study at hand also examines if ethical concern and socio-demographic characteristics of a consumer are drivers of the importance he/she attaches to the fair trade label. Prior literature finds contradicting results on this issue. We conclude that it is not possible to predict sensitivity to a fair trade label by socio-demographic variables, while it can be done by ethical concern.

(7)

of countries. Yet, much of the world’s population remains poor and the gap in living standards between the rich and poor needs to be narrowed substantially, says the United Nations (2010). Consequently, if the development transition were to follow the same consumption and production patterns, pressures on life-support systems would become intolerable. However, the world population is expected to be 9 billion in 2050, there is a rise in global affluence and associated consumption and this is expected to put an even higher and increasing pressure on resources (United Nations 2010, WWF 2006). Furthermore, there are increasing discussions worldwide on the present consumption society and there is a growing conviction that consumption needs to be restrained in prosperous economies. These issues add urgency to the study of ethical consumption (Newholm and Shaw 2007). Moreover, because changes are needed in consumer lifestyles and in the way consumers choose and use products to bring consumption to a sustainable level (World Business Council for Sustainable Development 2008), understanding buying behaviour of consumers is critical. Additionally, especially consumers in developed countries like Holland need to take the lead in moving towards sustainable patterns of consumption (United Nations 2010). The ambition of the Dutch ministry for Agriculture, Nature and Food Quality is to be in the international lead of production and consumption of sustainable food in 2015 (Ministry LNV 2009). This study contributes to a better understanding of sustainable buying behaviour in the Dutch market and this is a small step towards reaching the goal of the Ministry.

1.3 Structure of thesis

(8)

2. MARKET BACKGROUND

This chapter gives a background of sustainable products in general, and in specific on fair trade products on the Dutch market.

2.1 Definition of sustainable products

Sustainable products are products produced and processed while taking care of people or the environment, either close to home (e.g. healthy food products, organic food, animal-well-being) or faraway in the world (e.g. fair trade products, legally logged wood) (DePelsmacker et al. 2005, Ministry of EL&I 2011). They concern issues as nature and environment, animal wellbeing, health, waste of food, social aspects and justice, more than is legally required (Ministry EL&I 2011). Sustainable products often contain a sustainable label. These labels are present in an increasing diversity (Connolly & Shaw 2006). As such, present on the Dutch market are amongst others the Eco label for ecological products, Marine Stewardship Council (MSC) label for sustainable fish and the Max Havelaar label for fair trade products. Most of the labels are licensed by labelling organisations and can be placed upon a product if it meets the specific conditions and rules set by the labelling organisation.

2.2 Growth of the market

The assortment of sustainable products is growing fast these recent years and sustainability is becoming more visible for the consumer in retail stores (Bramer 2011, Milieudefensie 2011, NCDO 2011). In this line, the assortment of the two major product types in sustainable products, ecological products and fair trade products, grew with respectively 30% and 40% in 2011 (Ministry EL&I 2011). As a result, the average number of organic (116) and fair trade products (28) in Dutch supermarkets is now more than ever before (Milieudefensie 2011). Furthermore, A-brands have started to produce sustainable and several products of A-brands are carrying a sustainable label (for example Verkade and Van Gilse carry the Max Havelaar label). Also retail chains have launched their own private labels with sustainable products (such as Puur&Eerlijk by Albert Heijn and Fairglobe by Lidl).

(9)

and smaller labels which account for less than 3% of expenses such as free range chicken and eggs (Ministry EL&I 2011).

2.3 Max Havelaar fair trade products

The Max Havelaar label guarantees that the products are produced fair conditions (concerning working conditions, living conditions and child labour) and that farmers receive a minimum income. More information about the Max Havelaar label can be find in Appendix A. In the Dutch market, a wide spectrum of food products carry the Max Havelaar fair trade label: fruit, vegetables, rice, herbs, sugar, honey, oil, sandwich filling, thee, coffee, beer/wine, fruit juice, chocolate, ice cream, candy, cookies and snacks. The supermarket is the most important distribution channel where most consumers buy their fair trade products (NCDO 2011). The turnover of Max Havelaar certified food products increased spectacular with 155% from 2009 to 2010 (Ministry EL&I 2011). Also, the retail worth of certified fair trade products grew from €87 million in 2009 to €120 million in 2010 in Holland. The sales in kilos/litres was €23 million in 2010 with an increase of 57% with respect to 2009 (Max Havelaar, 2011b). This is divided as the following in different product categories: chocolate products has the most buyers, followed by fruits and third is coffee. Furthermore, 49,5% of Dutch households bought one or more fair trade food products in 2010, which is 3,6 million people, with respective to 42,3% in 2009 and 29,8% in 2008. Moreover, fair trade buyers have become ‘heavier buyers’; the quantities, amount of purchases and the amount of money spent per household has increased (NCDO 2011).

(10)

3. THEORETICAL FRAMEWORK

This chapter covers relevant literature in the proposed area of interest. First, we discuss the importance of price, functionality and sustainable attributes of a product in consumer decision making. Following, the role of ethical concern and socio-demographic variables of consumers in the purchase of sustainable products is explained. Then, it is reported to what extent habitual behaviour and the environment of consumers relate to the choice for a sustainable product. Finally, we present a conceptual model.

In our empirical study, we focus on fair trade ethical features by referring to a fair trade label that is used on products in the Dutch market. We study the Max Havelaar label because it is the largest fair trade brand in Holland and has a high brand awareness (Max Havelaar 2011a). However, we also consider relevant findings on other sustainable products and attributes in our theoretical framework. More studies have examined these attributes simultaneously (e.g. Auger et al. 2003, Auger et al. 2008, Luchs et al. 2010).

3.1 Importances of product attributes

(11)

Studies in sustainable buying behaviour have found that consumers value the diverse attributes of a food product in a different way. First, DePelsmacker et al. (2005) studied the value/importance of different product attributes in buying behaviour of coffee, including a fair trade label. The authors cluster their respondents on the basis of the importance they attach to the different product attributes and they found four different clusters. Overall, the authors found that the brand attribute has the highest relative importance in purchase decisions, following by the flavour of the coffee and the fair trade label is third. Likewise, Tagbata and Sirieix (2008) find three consumer segments based on their willingness to pay (WTP) for fair trade and organic chocolates. They find that a substantial amount of their respondents, more than half, were not sensitive to organic and fair trade labels in a product; these consumers attach the most importance to price. Overall, they find that taste and health concerns overrule ethical arguments. Then, in a similar research, Auger et al. (2009) find three different segments based on consumers their preferences for products with social attributes; a brand, price and social conscious segment, in which the consumers placed greater importance on these attributes than respondents in the other two segments (Devinney et al. 2010). To sum up, following the findings of Auger et al. (2009), DePelsmacker et al. (2005) and Tagbata and Sirieix (2008), we expect to find that consumers attach different importances to the diverse attributes and expect to find different segments based on these.

It is argued that to purchase or not to purchase a fair trade product can be seen as deciding between two alternative actions, one more ethical than the other; which is some sort of social dilemma (Nicholls and Lee 2006, Van Doorn and Verhoef 2011). The consumer will evaluate individual motives that are more beneficial to themselves as an individual (e.g. functionality of a product or cheaper products) alongside ethical motives which have an ethical benefit (e.g. a better environment). In the next subsections a trade off between functionality and sustainability and price and sustainability is discussed.

3.1.1 Trade-off between functionality and sustainability

(12)

a satisfactory functional performance. Therefore, the consumer would choose for functional performance, which is in line with the principle of functional precedence that favours functional attributes over other attributes. Yet, once functional performance of a product is on a satisfactory level, feelings of distress decrease and consumers would choose for sustainability to minimize guilt. They find support for this in their empirical study. Similar, Auger et al. (2008) find that a ‘consumer will choose the product with positive social features when there is no sacrifice of functionality, but any hint of poor functionality will cause demand to collapse’ (Devinney et al. 2010: 97). Auger et al. (2009) conclude from this that consumers are not willing to sacrifice (good) functionality for social desirability (accepting a bad functionality). They find that this applies to all three of their found segments, including the social conscious segment. Moreover, Auger et al. (2008) argue that good social attributes cannot compensate for weak functional attributes; purchase intentions decrease in the case of ‘bad’ functional attributes, even if the product did have good social attributes. Additionally, Auger et al. (2008) find that consumers will switch to a product with negative ethical features to obtain good functional features. However, this does not mean that products are not harmed by negative ethical features; unethical behaviour of an organization affects the attitudes of consumers towards an organization, even when the product has superior functional features (Folkes and Kamins 1999).

(13)

Hypothesis 1. Consumers are not willing to trade off their preferred level of functionality for sustainability.

3.1.2 Trade-off between price and sustainability

As argued in the previous section, it seems that consumers are not willing to sacrifice a good functionality for sustainability. Yet, what about price; will consumers trade off a lower price (accepting to a higher price) for sustainability? And are they willing to pay more for a product once it attains sustainable attributes? This is an important question for the profitability of sustainable products as the price for sustainable products is generally higher than for their non-sustainable equivalents. Nevertheless, the literature is not clear on these issues.

(14)

al. (2008) find all of their respondents reacting to price as expected; as price rises, the likelihood of purchasing a product with social features declines.

Following these findings above, we expect that there exists a (small) willingness to pay more for a product with a fair trade label. However, this stops at a certain price, as consumers do not seem willing to pay the actual price premium. Following this, we suppose that an interaction effect exists between price and a fair trade label. For example: consumers value a product with a fair trade label more than a product without a fair trade label at a low price, while this less the case at a higher price. We propose the following hypothesis:

Hypothesis 2. There exists an interaction effect between price and a fair trade label.

3.2 Segmenting the ethical consumer

3.2.1 Socio-demographic variables

Auger et al. (2003: 296) argue that socio-demographic variables can be used for ‘identifying consumers who are more likely to respond to an ethical proposition’. In the next sections, we discuss the role of education, income, age, gender and living district.

Education, income: McGoldrick and Freestone (2008) find that education is modestly

significantly associated with the willingness to pay ethical product premiums. Furthermore, DePelsmacker et al. (2005) found that highly educated respondents (more than 12 years of education) attach more importance to the fair trade label than to other product attributes. In line with this, the most important difference between fair trade buyers and non fair trade buyers in the Netherlands (someone who bought something in the recent half year is a buyer), is that buyers are significantly higher educated (minimal HBO level) (NCDO 2011). Moreover, NCDO (2011) concludes that the higher educated households and/or households with relatively higher incomes are more open to fair trade products. Drawing on these findings, it seems likely that consumers that are higher educated and/or have a higher income will attach more importance to the fair trade label in a purchase decision.

Age: DePelsmacker et al. (2005) found that respondents who attach more importance to the

(15)

likely that middle-aged and older consumers attach more importance to the fair trade label in a purchase decision.

Gender: Vermeir and Verbeke (2006: 171) state that prior literature indicates that ‘gender

does not seem to influence ethical decision making’. Carrigan and Attalla (2001) find the same result. Additionally, DePelsmacker et al. (2005) find that respondents that differ in their attached importance to the fair trade label do not vary in gender. Thus, it seems that there is no relationship between gender and sustainable buying behaviour.

Living district: NCDO (2011) studied if there is a relation between the living district of

households in the Netherlands and their purchase behaviour of fair trade products. They found that households in the North and South of the Netherlands (areas with less population density) buy relatively less fair trade products.

Opposing to these findings, there is also a body of research arguing that socio-demographic variables do not influence sustainable buying behaviour (McGoldrick and Freestone 2008, Roberts 1996). Additionally, Auger et al. (2003, 2008, 2009) find that socio-demographics ‘play no significant or predictable part in either the choice of product or the degree to which social features matter’ (Devinney et al. 2010: 106) and have no relation to the willingness to pay for the sustainable product (Auger et al. 2003). Furthermore, Auger et al. (2009) do not find differences in socio-demographics between their found segments based on sustainable buying behaviour. Scholars have argued that the finding that demographics have no influence can be explained by the fact that the market is extending, the products have become widely available and that ethical concern, responsibility, awareness and consumerism have become prevalent and the norm among all segments of the society (Carrington et al. 2010, Doran 2009, Roberts 1996).

In conclusion, findings in the literature seem to be contradicting. Therefore, it is interesting to examine if socio-demographic variables of consumers are a driver of the importance a consumer attaches to a fair trade label. Following the findings of Auger et al. (2003, 2009), Devinney et al. (2006, 2010) and McGoldrick and Freestone (2008), we purpose the following hypothesis:

(16)

3.2.2 Personal values & ethical concern

NCDO (2011) asked Dutch consumers who bought fair trade food products what their reasons were to buy these (apart from coffee) and they mostly specified an idealistic reason such as ‘to support the famers in poor countries’ (NCDO 2011). Following this, it seems that consumers’ their ethical concerns can be related to the probability of buying a sustainable product. In this line, multiple scholars have tried to define and segment the ethical consumer, for example by using personal values (DePelsmacker et al. 2005, Doran 2009), surveys with questions regarding ethical stances (Auger et al. 2003, 2008), asking the consumer to rate ethical issues (Luchs et al. 2010) and manipulating the consumer about their involvement with sustainability (Vermeir and Verbeke 2006). We discuss some results below.

First, Vermeir and Verbeke (2006) find in their study that the attitude towards buying sustainable dairy products (that correlates strongly with intention to buy) is significantly positively impacted by the involvement with sustainability. Second, DePelsmacker et al. (2005) argue that people’s values appear to have a significant impact on their ethical consumption behaviour. The authors use the Rokeach Value Survey (Rokeach 1973) and find that respondents attaching the most importance to a fair trade label of all attributes under study are less conventional and more idealistic than respondents who attach the most importance to the brand attribute and flavour attribute. In the same line, Doran (2009) find a relation between personal values and purchase of fair trade products, using the values of Schwartz (1992). The author argues that consumers who start to consume more fair trade, will be less involved with in-group (benevolence), yet more involved on all people and nature (universalism). Summarizing, these findings seem to indicate that (idealistic) values and ethical concerns of a consumer play an important role in sustainable buying behaviour. We purpose the following hypothesis:

Hypothesis 4. Ethical concern is a driver of the importance a consumer attaches to the fair trade label.

(17)

Following this, Devinney et al. (2010) argue that there are no covariates that matter and the heterogeneity lies deeper within the individual. Thus, prior findings on if ethical concern is a driver of the importance a consumers will attach to the fair trade label seem to be contradicting and therefore it is interesting to examine.

Despite the arguments before that it is not possible to segment sustainable buying behaviour on consumer characteristics, it is possible to segment consumers on their willingness to incorporate social features (Devinney et al. 2010). Yet, these segments are latent. In this way, as mentioned in the beginning of this chapter, DePelsmacker et al. (2005) and Tagbata and Sirieix (2008) find a social conscious segment based on the WTP for fair trade food products. Also Auger et al. (2009) find a social conscious segment based on consumers their preferences for products with social attribute. Therefore, we expect to find a latent social segment of consumers who attach more importance to the fair trade label compared to other latent segments.

3.3 Habitual behaviour

Habits are learned behaviour patterns that have led to satisfactory results in the past (NCDO 2011). Switching to a sustainable product requires that consumers depart from their usual product and brand and possible deep-routed habitual behaviour. Backus et al. (2011) conclude that habitual behaviour can predict fair trade food consumption. Furthermore, the authors find that habitual behaviour influences the relation between intention and behaviour of fair trade food products; logically, consumers who already buy fair trade products by routine will sooner translate their intention to purchase (Backus et al. 2011). Following this, it seems that habitual behaviour of other products can obstacle switching to ethical products (Backus et al. 2011). This can be a reason why ethical intentions are not always matched by purchase outcomes (McGoldrick and Freestone 2008); the ‘attitude-behaviour gap’.

3.4 Role of the environment

(18)

Fair trade label Functional attribute

Socio-demographic variables

Choice for a fair trade food

product

Interaction between price and fair trade label

Ethical concern +

~

+

*

~

means that we expect no effect.

First, for incorporating the social environment, Backus et al. (2011) adds identity, group norms and group pressure. The authors find that these constructs predict the intention and the perceived purchase of fair trade food products. The construct group norm has the strongest influence, which indicates that it is important that respondents see that people in their environment buy fair trade food. This is supported by the findings of Vermeir and Verbeke (2006) who find that intentions to buy sustainable dairy products can be explained by consumers experiencing social pressure from peers.

Second, Backus et al. (2011) measure the physical environment by the (perceived) availability of sustainable products. They find that availability predicts fair trade food consumption and also positively influences the relation between intention and behaviour of fair trade food products (Backus et al. 2011). This is supported by a study of NCDO (2011) on the Dutch market, in which consumers have indicated that a reason for not buying fair trade is that they do not encounter these products. Furthermore, it is found that consumers assess the availability of environmental friendly and ‘fair’ food lower than the availability of healthy food products (Milieudefensie 2011). This could be a reason of why intentions to buy remain low while attitudes are positive (attitude-behaviour gap) (Vermeir and Verbeke 2006). Following this, Vermeir and Verbeke (2006) show that ethical food consumption can be stimulated by increasing the perceived availability of consumers.

3.5 Conceptual model

Figure 1 presents the conceptual model as outlined in the previous sections, presenting the purposed relations.

FIGURE 1: Conceptual model

(19)

4. RESEARCH METHODOLOGY

In this chapter we discuss the research methodology of our study. First, we explain the use of a choice based conjoint analysis. Following, the product under study and the study design of the choice-based conjoint analysis is described. Next, we discuss the ethical concern and socio-demographic variables. After, the experimental procedure is explained. Last, we propose the plan of analysis.

4.1 Choice-based conjoint analysis

(20)

attribute and a close-to reality setting (DePelsmacker et al. 2005). Furthermore, the last years CBC has developed into the most widely used conjoint-related technique (Sawtooth 2009a). In the CBC study, respondents express their preferences by choosing combinations of attributes that compose products from a set of products and are required to make trade-offs. CBC is more realistic than other measures, as it is similar to what consumers actually do in the marketplace (DePelsmacker et al. 2005, Green, et al. 2001, Green and Srinivasan 1978, Sawtooth 2008). The CBC study enables to calculate utilities for each level of each attribute, that are consistent with the preferences of the respondent (Green and Srinivasan 1990). Hair et al. (2010) describe utility as ‘a selective judgement of preference unique to each individual’ (Hair et al. 2010: 66). These utilities are defined as part-worth’s and the overall utility of a product is received by summing up the part-worth’s of its attribute levels.

The methodology of CBC also has some drawbacks. First, even though the choices in the CBC study are a realistic proxy of buying behaviour, not all respondents are potential buyers: they may not be interested or able to buy in the real market place. This is also argued by Carrington et al. (2010), who state that the assumption that intentions lead to behaviour (‘intention-behaviour assumption’) is empirically found as not correct in prior literature while most studies on ethical consumer behaviour focus on the relation between attitudes and intentions and assume that intentions lead to behaviour directly. Therefore, one should be careful in interpreting intention-based models of ethical behaviour (Carrington et al. 2010). Furthermore, conjoint analysis assumes that a respondent has perfect information and there is an equal availability of products. It cannot account for differences in awareness, knowledge of respondents and differences in availability of products. Moreover, the many attributes, levels and choice questions require the respondent to read and process a lot of information and this makes it an less efficient way to elicit preferences. Additionally, each choice does not show the relative preferences for the not-chosen alternatives and how much more their chosen option is preferred over the other options (Orme 2010, Sawtooth 2009a). Last, the importances and utilities of the attributes are relative, they only state the importance of this attribute on choice relating to the use of the particular attributes in the study (Malhotra 2007).

4.2 Product under study

(21)

vices’ (Van Doorn and Verhoef 2011: 2). Van Doorn and Verhoef (2011) found a negative quality inference for vice food with an organic label. However, taste and quality of the product are found to be motives for Dutch consumers to buy fair trade food products (NCDO 2011) and a large part of the fair food products bought in the Netherlands in 2010 are vice products (coffee, chocolate). This is probably due to a different perception of consumers of fair trade versus organic vice food products. If we examine a virtue product, according to Van Doorn and Verhoef (2011), we do not have to consider quality inferences of the label on the product. Yet, it seems that consumers only value pro-social benefits in combination with vice products (Van Doorn and Verhoef 2011). Furthermore, the literature regarding fair trade food almost always studies a vice product (Auger et al. 2008: athletic shoes, soap; Auger et al. 2009: athletic shoes, batteries, DePelsmacker et al. 2005: coffee; Luchs et al. 2010: shoes and mobile phone; Tagbata and Sirieix 2008: chocolates).

Orange juice contains both virtue and vice associations; it can be seen as a vice product because it contains more sugar and calories than its equivalent products, while it can be seen as a virtue product because it contains vitamins (Van Doorn and Verhoef 2011). Following this, we decided to study the product orange juice. The negative quality inferences of the label on the product will be minimized as it is not entirely a vice product, while we still can use the findings in the literature of fair trade vice products in our research. Additionally, we expect almost all consumers to be familiar with this product and expect a lot of consumers to buy this product on a regular basis. However, we still need to be aware of possible negative quality inferences of the fair trade label on the product.

4.3 Study design of CBC

4.3.1 Number of attributes and levels

(22)

levels: contains fruity bits or a smooth style; 3) The attribute taste has three levels: tangy/sharp/slightly bitter, slightly sweet or sweet. Second, we include four levels for our price attribute, to balance the number of levels across attributes in order to minimize the ‘number of levels effect’ (Johnson and Orme 2009, Orme 2003, Sawtooth 2009b, Wittink et al. 1992). Considering the prices of orange juices in the market at the moment of data collection (see Appendix B), we range our price levels between €0.95, in the range of price discounters and €2.99, in the range of freshly squeezed products of private labels. Then, the second and third price level are in the middle between these: €1,61 and €2,3. Third, the sustainable attribute is specified on two levels: a Max Havelaar fair trade label present or not present.

Summarizing, Table 1 presents the attributes and its levels. These attributes make up less than six attributes which follows recommendations in order to minimize fatigue and boredom of our respondents (Green and Srinivasan 1990, Orme 2002, Sawtooth 2009a, Wieringa 2009).

TABLE 1: Attributes and their levels

Attribute Level 1 Level 2 Level 3 Level 4

Type of juice Freshly squeezed juice Juice made from concentrate

- -

Texture Contains fruity bits Smooth style - -

Taste Tangy, sharp, slightly bitter

Slightly sweet Sweet -

Price €0,95 per litre €1,61 per litre €2,30 per litre €2,99 per litre Fair trade label Max Havelaar fair

trade label

No Max Havelaar fair trade label

- -

4.3.2 Design of choice task

Stimulus format: We use a full profile design instead of a partial profile design because we do

not have a large number of attributes. This means that we display a level from every attribute in every product profile in the study.

Number of profiles per task and the number of choice tasks: Increasing the number of tasks

(23)

difficult. Next, too many choice tasks can bias results and decrease the quality of the information. As a solution, a higher sample size makes it appropriate to use fewer tasks. Around 3-5 concepts per task is recommended, particularly if the attributes also have around three to five levels (Johnson and Orme 1996, Pinell 2005, Sawtooth 2009a). We tested multiple designs with a different number of tasks and profiles in a provided tool by Sawtooth (see Appendix C.1). In the test results we see that using 4 or 5 concepts does not change the minimum required number of respondents (150) and the number of tasks (10) in order to have an efficient design, while having 3 concepts does. Therefore, we decided to have a design of 10 tasks and 4 concepts, assuming we are able to obtain a minimum of 150 respondents. We also include three hold out tasks, which we explain further on in this chapter, which makes the total number of tasks 13. For this design we employed another test provided by Sawtooth and conclude that this is an efficient design (see Appendix C.2)

Dual None option: A choice-based conjoint study can incorporate a none option which

enables respondents to chose none of the alternatives. However, Sawtooth (2009a) note that as the number of attributes increases, the propensity to choose the none option also increases. Furthermore, there is no information on the attractiveness of profiles if the none option is selected (Brazell et al. 2006). As a solution, we employ a dual none option as a second-stage question: respondents are first asked to choose among alternatives and after this, they are asked if they would really buy the alternative they selected. It is discussed to be more realistic and to reflect actual purchase intentions. Furthermore, it is more pleasant because respondents do not have to choose products that they do not like (Johnson and Orme 2003, Sawtooth 2009a).

4.3.3 Design generation method

(24)

3.2.4 Hold out tasks

We include three hold out choice tasks in our study. These tasks are fixed; all respondents receive exactly the same product combinations in these tasks (Sawtooth 2009a). These are not used for part-worth estimation but employed to check validity of estimated utilities, estimated by the ability of the model to predict choices of the hold out task. Herewith, we can compare different models by their predictive accuracy (Johnson and Orme 2009, Sawtooth 2009a). In designing hold out tasks, it is important that the profiles are not equally attractive or only one of the choices is preferred (Johnson and Orme 2009, Sawtooth 2009a). Furthermore, Johnson and Orme (2009) recommend to have level overlap in the hold out tasks. Next, Johnson and Orme (2009) argue that hold-out tasks should be spaced evenly throughout the choice questionnaire for a reliable indication of validity. Therefore, we placed our hold-out tasks in the 3rd, 7th and 11th positions (see Appendix D for the designed hold out tasks).

4.4 Ethical concern and socio-demographic variables

(25)

TABLE 2: Included variables

Variable Item Level Scale

Ethical concern

People should be willing to help others who are less fortunate ordinal 7-point Likert Scale

Helping troubled people with their problems is very important to me People should be more charitable towards others in society

People in need should receive support from others

Age Age of the respondent in years ratio absolute number

Gender Gender of the respondent nominal 2 categories

Income Netto income per month of the household of the respondent ordinal 10 categories Education

level Highest level of education completed by the respondent ordinal 7 categories Living district District in the Netherlands where respondent lives nominal 5 categories Household

size Size of the household of the respondent ratio absolute number

Working Number of persons working in the household of the respondent ratio absolute number Working

fulltime

Number of persons working more than 20 hours per week in the

household of the respondent ratio absolute number

Children Amount of children that the respondent has ratio absolute number Children at

home Amount of children that lives in the household of the respondent ratio absolute number Frequency of

consumption

Number of times the respondent drinks orange juice bought by the

household ordinal 9 categories

4.5 Experimental procedure

(26)

4.6 Plan of analysis

Data from a choice-based conjoint study is traditionally analysed at the aggregate level. Aggregate models assume that variability among the respondents is random. However, scholars argued that consumers are unique and have unique preferences and the aggregate model fails to take this into account. Therefore, aggregate models result in poor share predictions and are not appropriate. Furthermore, it is argued that aggregate models endure from the Independence from Irrelevant Alternatives assumption (IIA) (Howell 2009, Magidson et al. 2003, Orme 2009b, Sawtooth 2004). There are two other methods available: the Latent Class analysis (LC) and Hierarchical Bayes’ (HB) estimation which both account for consumer heterogeneity. We discuss these methods in more detail below.

4.6.1 Hierarchical Bayes’ modelling

While we have limited data at the individual level (only 10 choice tasks), we still want to estimate individual level utilities. This can be overcome by (hierarchical) Bayesian methods which have become widespread in Marketing. We explain the principals behind hierarchical bayesian modelling in this subsection.

Convential, non-bayesian analysis assumes that a model with particular parameters describes the data and it is checked whether the data is consistent with the assumptions, examining the probability distribution of the data given a hypothesis. The other way around, Bayesian analysis examines the probability distribution of particular hypothesis (H ), given the data. It uses conditional probability and uses the following rule:

where means ‘is proportional to’, and is a set of data.

(27)

form the posterior estimate in which the prior information is combined with information from the data. The HB model uses the above described bayesian way of updating probabilities to estimate parameters of the model; after each iteration the probabilities are updated.

The HB model contains two levels and is therefore named ‘hierarchical’. The higher level of the model assumes that the part-worth’s of respondents are explained by a multivariate normal distribution with a vector of means and a matrix of covariance’s:

where is a vector of part-worth’s for the respondent, is a vector of means of the distribution of respondent’s part-worth’s and is a matrix of variances and covariance’s of

the distribution of part-worth’s across respondents.

The lower level assumes that a multinomial logit model runs the probabilities of respondents of choosing particular alternatives. Hereby, the probability of the _{respondent choosing the}

_{alternative is:}

where is the probability of an respondent choosing the _{alternative in a choice task and} is a vector of values that describes the alternative in the choice task.

Thus, the HB model needs to estimate the parameters , and . By default, these parameters are set to zero by the Sawtooth HB program and we follow this in estimating our model. They are estimated by an iterative process and each iteration proceeds like following (Monte Carlo Markov Chain: MCMC): 1) first, present estimates of the betas and are used to form a new estimate of , 2) with use of the estimates of the betas and , we generate a new estimate of and 3) new estimates of the betas are draw, using estimates of and , with a procedure called ‘Metropolis Hastings Algorithm’.

(28)

from to and the speed of convergence. Sawtooth CBC/HB uses an adaptive simulation algorithm.

The likelihood of the data given each set of part-worth’s is computed with the logit formula described before. The probabilities of each choice that the respondent has made are calculated and then multiplied; resulting in values and . Furthermore, the relative density of the distribution of the betas corresponding to and is calculated given parameter estimates of and : resulting in and . The following formula is used:

Then, the ratio is calculated by the following formula:

This is ‘the ratio of posterior probabilities of those two estimates of the beta, given current estimates of and , as well as information from the data’ (Sawtooth 2007: 14). If the ratio is greater or equal to unity, we accept because it has a posterior probability that is greater than or equal to . Yet, if ratio is less than unity, a random process determines if will be accepted or rejected. This process continues until convergence; if there are no more increases possible in the betas and its fit to the data. Two stages can be identified in the iteration process of Sawtooth CBC/HB. In the first stage, the parameters are moving to convergence and these results are not saved; the ‘burn-in’ iterations. Then, in the second stage convergence is assumed and, individual-level utilities for each respondent are obtained by averaging the values of each iteration in this stage. This is to minimize the effect of initial values on the posterior inference.

(29)

4.6.2 Latent Class modelling

The LC method divides individuals into latent classes with similar preferences in the groups and different preferences between the groups, founded on the choices made in the CBC questions. It uses the logit rule, as explained before. Part-worth’s are estimated for each segment. Also, for every respondent a probability that he/she fits to all obtained segments is calculated. In comparison, other methodologies attribute each respondent exclusively to one segment. It argued that LC analysis provides better results than combining K-Means cluster analysis with an aggregate logit model if the data contains within-cluster heterogeneity. Furthermore, it is argued that the provided statistics with the analysis are superior than statistics of alternative cluster analyses (Sawtooth 2004, Sawtooth 2009a).

To sum up, LC models utilize a discrete distribution of heterogeneity as opposed to a continuous distribution in HB. HB models assume that each respondent has their own unique preferences, while LC models assume that each respondent belongs to some extent to a latent cluster, which has its own unique preferences (Magidson et al 2003). The HB estimation is argued to lead to more accurate predictions. Moreover, it is said to successfully reduce the IIA problem (better than LC models) and to improve the predictive validity of individual-level models and market simulation (Orme 2006, Orme 2009b, Sawtooth 2004, Sawtooth 2009a). The majority of CBC users employ this method and it is recommended by Sawtooth (Sawtooth 2009a) and Pinell (2005).

(30)

5. ANALYSIS AND RESULTS

In this chapter we report the analyses performed and the obtained results. We begin with a description of the obtained sample. Next, the estimation of the part-worth’s, first the HB model and then the LC model is discussed. Following, we describe the found importances of the attributes of the aggregate model and the latent classes found. Also, the price elasticity and WTP is examined. Then, we investigate if ethical concern and socio-demographic characteristics of a consumer are a driver of the importance consumers attach to the fair trade label. Last, we discuss the find results and relate them to prior literature and our defined hypotheses.

5.1 Sample description

The link of the survey was clicked on 469 times after which the survey was abandoned during the introduction screen 55 times. Of the 414 surveys left, 128 surveys (30,9%) were incomplete and abandoned somewhere during the survey. We have 286 completed surveys. Yet, we deleted nine surveys because the respondents were under 18 and/or are living at home with parents as these are argued to be no real consumers yet. Of all completed surveys, ten respondents finished in less than 4,5 minutes, which can be argued too short to have processed the choice tasks accurately (average response time: 9 minutes, 40 seconds). One respondent finished earlier than he started, and we assume that the data is not correct. Furthermore, we remove four respondents who took longer than one hour to complete the survey. This leaves us with 262 valid respondents. However, the maximum number of respondents we can analyse is 250 because of the academic license of the programme we use. As a solution, we removed twelve respondents either aged 24-25 and/or living in Groningen, because these categories are overrepresented which gives us a more evenly spread sample. This results in a sample of 250 respondents. Each choice (1-4) received around 25% of all choices, as would be expected (see Appendix F). This sample size is above the calculated minimum of 150 respondents for an efficient design of a CBC study (see section 4.3.2).

(31)

TABLE 3: Sample description Variable Categories % Age 20-30 years 34,8% 31-40 years 24,8% 41-50 years 21,6% 51-60 years 14,8% 61-70 years 4,0% Gender male 48,0% female 52,0% Education level

Little educated (elementary school, lower vocational technical school) 1,6% Intermediate educated (high school, intermediate vocational education) 11,2% High educated (pre-university education, higher vocational education,

university) 87,2% Living district

(Conurbaties of) Amsterdam, Rotterdam, The Hague 8,0% Remaining’s of North Holland, South Holland, Utrecht 13,6%

Groningen, Drenthe, Friesland 63,2%

Overijssel, Gelderland, Flevoland 11,6%

Zeeland, North Brabant, Limburg 3,6%

Income

Little to modal 20,4%

Above modal till 2x modal 40%

Above 2x modal 27,2%

Do not wish to say 12,4%

HH size 1 person 26,4% 2 persons 37,6% 3-4 persons 30% 5-6 persons 6% Frequency of consumption

Very often (multiple glasses per day to 2-6 glasses per week) 42% Often (1 glass per week or 1 glass per 2 weeks) 35,2%

Occasionally (2-11 glasses per month) 11,6%

Sporadically (1 glass per month) 5,2%

Never (1 glass per year or less to never) 6%

Ethical concern (Average answer) Strongly disagree: 4% Disagree: 8,4% Neutral: 16,8% Agree: 66,8% Strongly agree: 7,6%

(32)

increases the reliability of our results, because consumers that do not consume orange juice regularly would not be familiar with the buying situations that are presented to them in the survey.

5.2 Part-worth estimation

In this section we describe the modelling used to estimate the part-worth’s of our respondents. First, we report the Hierarchical Bayes’ modelling, followed by the Latent Class modelling. 5.2.1 Attribute relations

The attributes type, texture and fair trade label only consist of two levels, so they are estimated with a part-worth relation. Yet, the attributes taste and price have more than two levels and require a decision of the kind of part-worth relationship to be used (Hair et al. 2010). We expect that each respondent has a preference for a different kind of taste and therefore we expect a part-worth distribution. For the attribute price, we assume that each respondent has a preference for the lowest possible price and expect a (negative) linear distribution. We tested the relations of the attributes price and taste with a HB model.

5.2.2 Hierarchical Bayes’ modelling

We estimated the HB model with use of the software Sawtooth CBC/HB. Sawtooth uses as default a particular number of iterations before assuming convergence (initial burn in iterations) and 10.000 iterations after assuming convergence. Only these latter iterations are used to create point estimates of the parameters for each respondent; the point estimates of the part-worth’s of each respondent are equal to the average estimates of the individuals betas for the last iterations. From our literature framework, we expect there is an interaction effect present between price and the fair trade label. Yet, we are curious to see if the fair trade label also has interactions with the other attributes (type, texture, taste). Furthermore, we estimated an extra model with price coded as linear to test for the distribution of the attribute price. Concluding, we estimated four different models: 1) a model without interactions, 2) a model with the following interactions: fair trade label x price, fair trade label x type, fair trade label x texture and fair trade label x taste, 3) a model with an interaction for fair trade label x price and 4) a model with the attribute price coded as linear.

5.2.2.1 Convergence of the models

(33)

the Markov chain is converged to its stationary distribution and therefore, we should be cautious. Also, after convergence, if variance in the estimation becomes stable, a considerable amount of variation will still be in the parameter estimates (Allenby and Rossi 2006, Allenby et al. 2005, SAS Institute Inc. 2009, Sawtooth et al. 2009b). We analysed the convergence of the four models estimated with initial iterations of 10.000, 20.000 or 30.00 by 1) visual analysis through trace plots, 2) statistics that indicate goodness of fit and 3) statistical test diagnostics Gelman and Rubin diagnostic and Geweke diagnostic (see Appendix H.1, H2 and H.3). Deriving from this, we can assume that all four models have reached convergence in the first 10.000 iterations.

5.2.2.2 Validation of the models

Model fit statistics: The CBC HB analysis gives four statistics that indicate the goodness of fit

of the model (see Appendix H.2). First, we see that the model with multiple interactions has the highest Pct. Cert. with a value of 71,7% which means the log likelihood is 71,7% on the way from chance to perfect fit; thus the model is better than chance. Second, for RHL we find the highest value of 0,641 for the model with multiple interactions. This is 2,5 times better than the chance level (which would be ¼ by using 4 profiles). The last two statistics, ‘average variance’ (avg. variance) and ‘Parameter RMS’ (PRMS) are indirect indicators of goodness of fit. These statistics increase in the first iterations for all models (see Appendix H.1 and H2) which suits a good model. Following these statistics, the model with multiple interactions seems to have the best fit.

Predictive validity: To assess the predictive validity and reliability of our models, we

(34)

However, it correspondents with the number of alternatives; more alternatives, a lower MAE. Therefore, we also estimated the mean squared error (MSE), which is the average of the squared error. Following these findings, The model with multiple interactions has the lowest values for MAE and MSE, which indicates this model has the best fit (Malhotra 2007, Orme 2006, Orme and King 1998).

5.2.2.3 Testing attribute relations

It seems that attribute price has a near linear relation because the aggregated utility shows a negative linear slope. However, the difference between utilities is not equal to the difference in price levels. Also, Pinell (2005) argues that price should be coded as part-worth instead of linear, because it provides better ‘match in-market’ pricing changes and the notion of psychological pricing. Therefore, we code this attribute as part-worth, despite the higher predictive accuracy of the model with price coded linear compared to the model with price coded part-worth (Malhotra 2007) (see Appendix H.6). Also, this could lower the degrees of freedom and the statistical efficiency of the model (Malhotra 2007, Sawtooth 2007). Next, the attribute taste shows us that the level slightly sweet has the highest utility. However, there is also a fair group of respondents who prefer one of these other two tastes (see Appendix J.2) and therefore we code this attribute as part-worth.

5.2.2.4 Testing interaction effects

(35)

Concluding, we use the utilities of the model with an interaction between fair trade label and price because it has the highest hit rate and good statistics of model fit. Nevertheless, all of the above estimated models provide us with almost similar relative importances of attributes and an equal order of importance of the attributes (aggregated over all respondents).

5.2.2.5 Reversals in individual utilities

Most respondents have the highest utility for the lowest price and most respondents’ utilities decrease if price increases. However, this is not true for all respondents; some respondents prefer a higher price over a lower price. These are called reversals; patters of part-worth’s that are theoretically inconsistent. Indeed, there are 55 respondents with a greater utility for the second, third or fourth highest price than for a lower price. These can exist because of random noise in the data and/or can be the result of price insensitive consumers (Orme 2010). A solution would be to eliminate these respondents or to apply utility constraints in our estimation. This latter method has drawbacks (Orme 2010). However, because the difference between part-worth’s is small and the predictive accuracy for these respondents is good, we keep these respondents in our sample (Hair et al. 2010).

5.2.3 Latent Class model

We estimated LC models using the software Sawtooth Latent Class. We examined the distributions of the attributes earlier (in section 5.2.2) and therefore, we estimate all attributes with a part-worth relation. However, we also estimate a model with price coded as linear to be able compare these models. Next, in the HB model we discovered an interaction effect between fair trade label and price. Nevertheless, we also estimate a LC model with and without this effect and compare these models. Thus, we estimated three different models: 1) a model without interactions, 2) a model with a price x fair trade label interaction and 3) a model with price coded as linear and a price x fair trade label interaction. Orme (2006) argues that if one wants to compare groups of respondents and discover significant differences, one needs a sample size of around 200 per (hypothesized) group. However, we did not obtain such a large sample size. Therefore, we need to be cautious with the obtained results of the LC analysis.

(36)

differ significantly on ethical concern and socio-demographic variables by means of chi square tests (see later in this chapter).

5.2.3.1 Validation of the models

Model fit statistics: Sawtooth LC provides us with four statistics of model fit: Pct. Cert.,

Consistent akaike information criterion (CAIC), Chi square (Chi sq.), relative chi square (rel. chi sq.) (see Appendix I.1). First, we see that fitting a second group to our data means a relative gain, because the values of Pct. Cert and Chi sq. are higher for the 2 group solution than for the 1 group solution across all models. Second, the model with the interaction effect has higher values for Pct. Cert, which means this model is slightly better than the model without the interaction effect, in comparison to a chance model. Third, the Pct. Cert. of the model with price coded as linear is smaller than for the model with price coded as part-worth at all group solutions. Therefore, we decided to use the model with price coded as part-worth and an interaction between price and fair trade label; it has the best fit.

Number of segments: We see the values of Pct. Cert. and Chi sq. increase at the number of

groups, however not so much more after 4-group solution. The same applies to CAIC, which stops dropping greatly after the 4 group solution. Furthermore, the CAIC value is smallest for a 5-group solution (see Appendix I.1). Because the goal of the LC model is to come up with meaningful segments instead of developing an accurate simulator, we choose a solution based on interpretability and stability (Sawtooth 2004). Following this, the solution with 5 groups and also with 4 groups provides us with small sized groups and this is not desirable (see Appendix I.2). This leads us to choose the 3 group solution. We rerun the model again with a different starting point and we obtained the same results. This indicates that it is a stable solution (see Appendix I.3).

(37)

Predictive validity: Each respondent received probabilities of membership to each segment.

However, in calculating the predictive validity, we assume a respondent belongs to the segment for which he/she has the highest probability. We find a hit rate of 60%, a MAE of 6,1% and a MSE of 0,78%. This indicates a good model fit and that the model is better than chance (hit rate would be 25%) (see Appendix I.6). Moreover, we can also calculate ‘pseudo-individual utilities’ with the ‘pseudo-individual probability of segment membership and the utilities of each segment (Sawtooth 2004). With these pseudo-individual utilities, we find an average hit rate of 75,5%, a MAE of 6,4% and a MSE of 0,76% (see Appendix I.6). These are higher than when classifying a respondent in a distinct segment.

5.2.4 Comparing LC and HB utilities

In comparison, the HB model has a better predictive validity than the LC model, also when using the pseudo-individual utilities. Apparently, the HB model does a better job in predicting individual responses. This corresponds with the literature which argues that HB estimations lead to better predictions (Orme 2006, Orme 2009b, Sawtooth 2004, Sawtooth 2009a). However, we do find a correlation coefficient of 0,815 between the pseudo-individual utilities of the LC model and the individual utilities of the HB model.

5.2.5 Resulting utilities

(38)

TABLE 4: Obtained utilities

5.3 Utilities of product attributes

In this section we discuss the importance of the attributes, by their relative importance and by a sensitivity analysis in which we examine the impact of a change in an attribute on share of preference. Also, we examine how respondents behave when faced with a trade-off between their preferred level of a product attribute and a fair trade label. Moreover, we investigate the willingness to pay for the fair trade label, the price elasticity and an interaction effect between price and the fair trade label.

Attribute Level Aggregated HB model

Fair trade conscious segment Smooth style lovers segment Price sensitive segment Standard error

_Utility _{average modus median} _Utility _Utility _Utility Type of

juice

Freshly squeezed

juice 1,656 0,756 0,667 0,743 0,836 0,587 0,642

Juice made from

concentrate -1,656 0,756 0,667 0,743 -0,835 -0,587 -0,642

Texture _{Contains fruity bits} _0,218 _0,757 _0,692 _0,732 _0,438 _-1,296 _0,213

_{Smooth style} _-0,218 _0,757 _0,692 _0,732 _-0,438 _1,296 _-0,213

Taste Tangy, sharp,

slightly bitter -0,100 0,814 0,774 0,813 0,069 -0,262 -0,114

_{Slightly sweet} _0,419 _0,738 _0,714 _0,731 _0,218 _0,042 _0,222

_Sweet _-0,319 _0,894 _0,956 _0,902 _-0,287 _0,220 _-0,108

Price _{€0,95 per litre} _2,761 _1,206 _1,537 _1,168 _0,426 _0,830 _2,469

_{€1,61 per litre} _1,276 _0,843 _0,831 _0,847 _0,337 _0,475 _0,899 _{€2,30 per litre} _-1,360 _0,939 _0,953 _0,931 _-0,251 _-0,464 _-1,007 _{€2,99 per litre} _-2,677 _1,215 _1,153 _1,194 _-0,512 _-0,841 _-2,362 Fair trade label

Fair trade label 0,590 0,602 0,578 0,597 0,373 0,273 0,148 No fair trade label -0,590 0,602 0,578 0,597 -0,373 -0,273 -0,148