The influence of brand type on the preferred agent in a service context: A Choice Based Conjoint Analysis

The influence of brand type on the preferred agent in a

service context: A Choice Based Conjoint Analysis

By

The influence of brand type on the preferred agent in a service

context: A Choice Based Conjoint Analysis

By:

Redmer (Nils) van der Veen 26-06-2017

MSc Marketing, Intelligence Master Thesis

University of Groningen Faculty of Economics and Business

Department of Marketing 9700 AV, Groningen

Supervisors: First: Dr. J. van Doorn Second: Dr. L. Lobschat

Student:

Redmer (Nils) van der Veen Address: Tuinbouwstraat 47-9

9717JB Groningen Phone: +31643754073

MANAGEMENT SUMMARY

Nowadays, the use of technology has increased in a service context. Over the past decades, the use of service agents has changed. The physical appearance of a service agent plays an important role. However, the influence of appearance on consumers’ preferences for a certain service agent is still not very well investigated. The aim of this study is to investigate to what extent the appearance has an influence on the preferred service agent. Moreover, this research will try to answer the question to what extent differences with regard to consumers’ preferences for a type of service agent could be explained by the influence of consumers’ tendency to anthropomorphize and brand type. Previous research focused on the robots’ appearances in different settings (E.g. health care, education). However, there has been no research conducted with regard to the influence of service robots’ appearances in combination with brand type and peoples’ tendency to anthropomorphize. The results are based on the input of 146 respondents that participated in this study. A Choice Based Conjoint analysis and a Latent Class analysis are estimated to check what type of service agent is preferred in the service frontline per segment. Besides, the respondents were exposed to a questionnaire about anthropomorphism and technology readiness.

In the past, it was assumed that service agents are human. The results of this study confirm this assumption and suggest that people prefer a human employee over a robot as service agent. However, interesting is that the results of this study suggest that consumers prefer a machine-like robot over a human-machine-like robot in general. There are two possible reasons for this. The first reason could be that people are already used to the appearance of a machine-like robot, since it is similar to some self-service technologies (SST’s). Another reason could be that consumers think the human-like robot is becoming too close to a real human being in appearance. Despite the expectations based on literature, the results of this study suggest that consumers do not consider service agents’ appearance as important as prior research claims. The relative importance of the appearance of the service agent is lower compared to the importance of type of brand and discount.

effect on consumers’ preferences for a human employee or human-like robot. Moreover, a premium brand has a positive effect on the preference for a human employee. A reason for this could be the ‘’personal touch’’ expected by customers from premium brands. Managers of premium brands should focus on human employees. People do not like a human-like or machine-like robot as a service agent. On the other hand, managers of economy brands should focus on human employee and although there is not significant evidence, it might be interesting to focus on the deployment of machine-like robots in the service frontline.

Finally, three of the four control variables (i.e. age, level of education, technology readiness), that were included are not significant in general. Only gender is significant in this study. However, exceptions can be made for specific relations between certain segments.

PREFACE

The passion for marketing started a couple of years ago. During my education on the community college, I became interested in the field of Marketing. The interest was strengthened during my internships at Hamilton Bright and Prodrivelearning.com. During my first internship, I did research on consumer behavior in association with the market leading brand of televisions at that time. I graduated after the second internship and decided to take a sabbatical. After nine months of working at Hamilton Bright and months of traveling through Asia, I started with the Pre-Master in September 2014. After a year of hard work, I started my Master Marketing Intelligence and now I am here, at the end of the Master Marketing. It was not always easy and I did not choose the path of the least resistance, but it was worth it. Writing my thesis is the last part of the journey as a student. During writing this thesis about the rise of the robots, I received feedback from my supervisor. Therefore, I would like to thank Ms. Jenny van Doorn for her help and great feedback during the period of writing my thesis. Secondly, I would like to thank Dr. Felix Eggers (assistant professor) for his valuable feedback on the methodology section. It was not always easy to write this thesis, but the first step is the hardest and finally I am done. Finally, I would like to finish off with a word of thanks to my family, friends and girlfriend who always supported me.

TABLE OF CONTENT

2.1 HUMAN-ROBOT INTERACTION (HRI) ... 12

2.1.1 Use of robots ... 12

2.1.2 Use of service robots ... 13

2.2 HUMAN-ROBOT INTERACTION (HRI) RESEARCH ... 15

3. CONCEPTUAL MODEL ... 15 3.1 CONSTRUCT ... 15 3.2 HYPOTHESES ... 16 3.2.1 Appearance Service agent ... 16 3.2.2 The influence of brand type ... 18 3.2.3 The tendency to anthropomorphize ... 20 3.3 CONTROL VARIABLES ... 21 4. METHODOLOGY ... 22 4.1 MEASUREMENT FOR CONSUMERS’ PREFERENCES ... 22 4.2 STUDY DESIGN ... 23 4.2.1. Attribute and attribute levels ... 23 4.2.2 Measurement propensity to anthropomorphize ... 25 4.2.3 Control variables - Measurement of control variables ... 26 4.3 CHOICE DESIGN ... 26 4.4 DATA COLLECTION ... 27 4.4.1 Procedure ... 27 5. RESULTS ... 29 5.1 DATA PREPARATION ... 29 5.2 DESCRIPTION SAMPLE ... 29

5.3 THE UTILITY, LOGIT & MNL MODEL ... 29

5.4 DIMENSION REDUCTION AND RELIABILITY TEST ... 30

5.5 AGGREGATE CHOICE-BASED CONJOINT ANALYSIS ... 31

1. INTRODUCTION

From 1750 till now there have occurred several industrial revolutions. First there was the rise of the steam train. After that, in 1862, there was the introduction of combustion engines. Later on, in the seventies of last century there was the invention of Internet and nowadays there is emerging a new industrial revolution: the rise of the robots. Like all the other revolutions there is media attention as well for this revolution (Dekker, 2017; Vries de, 2016; Markoff & Rosenberg, 2017). This implies that nowadays robots become more and more important in our society and it gives an overall view that the use of robots increases (Johnson, 2016; Meuter et al. 2005). The use of robots and devices can be in a professional setting or in a domestic setting. Danaher (2016) states that robots will take over a large amount of human activity, like car driving, cooking our meals, and delivering goods. In the past, robots were seen as subordinates to humans and there was a clear hierarchical structure when serving humans. The rapid change of technological developments ensures that the hierarchical relationship between humans and robots has changed. Nowadays robots are able to interact with humans (Kim, Park, Sundar, 2013).

perceptions of consumers (Butler & Agar, 2001; Canning, Donahue & Scheutz, 2014; Walters et al. 2008). The appearance of the robot plays an important role for companies since it influences consumers’ perceptions and expectations of the robot (Guo et al. 2017). Accordingly, the appearance of a robot can affect consumers’ perceptions regarding different aspects. In the field of health care, the appearance of robots can affect the perceptions of children with autism (Robins et al. 2004). Some robots are seen as more trustful or social because of their appearance (Li, Rau, Li, 2010; Broadbent et al. 2013). Koay et al. (2007) conducted a longitudinal research about the influence of appearance on perceptions and argued that preferences for a certain type of appearance changes over time. More important for this study will be the consumer perceptions about the competences of robots. Research suggest that appearance has an influence on people’s perceptions about the competences or abilities of the robot (Dautenhahn et al. 2005; Canning, Donahue & Scheutz, 2014; Scheutz, 2014). Appearance is a major driver of consumer perceptions and therefore this study will try to understand what type of appearance of the robot is most preferred in the service frontline. The literature about robots’ appearances is still not completely understood in a service context. Hence, this research is an expansion in de field of the existing robot literature.

with imitation of products. Hence, this study will take two different types of brand into account (I.e. premium vs. economy).

Another aspect that should be taken into account is the tendency to anthropomorphize. Van Doorn et al. (2017) argued that it is important to know when and to what extent people anthropomorphize technologies. The word is a combination from the Greek words; anthropos (human), and morphe (form or shape). Anthropomorphism is applicable to human interaction with various technologies (Epley, Waytz & Cacioppo, 2007). The level of peoples’ tendency to anthropomorphize is important for this research because it can influence the way of how people perceive brands, products, and services. From earlier research, it is known that marketers try to encourage the tendency of consumers to see products in a more relational way (E.g. name their cars) (Aggarwal & McGill, 2007). Understanding peoples’ differences in anthropomorphizing is important for identifying who is more likely to treat nonhuman-agents as humanlike (Waytz, Cacioppo & Epley, 2010). Additionally, understanding anthropomorphism is a critical factor for understanding when and how human interact with an increasingly wide variety of technological agents, like robots (Waytz, Cacioppo & Epley, 2010). Understanding individuals’ differences in the propensity to anthropomorphize is an important factor that will be taken into account in this research since consumers with a high level of tendency to anthropomorphize might be more likely to prefer a robot as a service agents. Therefore, the direction about the tendency to anthropomorphize is followed. The appearance of the agents will be measured on a scale of anthropomorphism, and therefore it might be assumed that there are differences between the agents in anthropomorphism. That is a different direction in the literature. Hence, the direction in literature about the individual differences for the tendency to anthropomorphize is more relevant for this research. That direction in the literature comprises the propensity to anthropomorphize as a consumer trait. Past research found that there are differences between individuals in the propensity to see humanity in objects (Waytz, Cacioppo & Epley, 2010). Accordingly, it is necessary to understand the insights of the process by which consumers attribute mental states and capacities to other entities (Haslam, 2006; Waytz, Cacioppo & Epley, 2010; Waytz et al. 2010). Hence, this suggest that is important when investigating the role and influence of the service agents’ appearance.

customers’ preferences for different types of robots used in the service frontline. The appearance of a robot is barely explored and can play an important role in the service frontline (Van Doorn et al. 2017). Another reason that emphasizes the importance of this study is to find out whether a marketing aspect, in this case brand type, plays a role in the context of service agents’ appearance (Eyssel & Kuchenbrandt, 2012). Hence, it needs to be examined to what extent brand type affect the preferred service agent in the frontline. Consequently, the literature abovementioned creates two research questions. The first research question is related to the impact of different types of robots on consumers’ preferences in the service frontline. The second research question is related to the different factors that might affect the relationship stated in the first research question.

1. To what extent does the appearance of service agents (human vs. robotic) have an

impact on consumers’ preference in the service frontline?

2. To what extent could differences with regard to consumers’ preferences for a type of

service agent (human vs. robotic) be explained by the influence of consumers’ tendency to anthropomorphize and type of brand?

The paper is structured as following. Firstly, this study will state an overview of the literature dealing with different types of appearance of robots, consumers’ allocated utility towards robots, and the factors that may affect this relationship. This literature overview purveys hypotheses and show hypothetical relationships. Based on existing literature there will be a conceptual framework created. In the second part, the methodology of choice-based conjoint analysis (CBC) is displayed and integrated with the different types of robots and brand name. Thirdly, the outcomes and results will be analyzed. Finally, the limitations, managerial implications, and the potential research directions are given.

2. THEORETICAL BACKGROUND

2.1 Human-robot interaction (HRI) 2.1.1 Use of robots

2.1.2 Use of service robots

Van Doorn et al. (2017) state that ‘’Automated Social Presence’’ (ASP) is the extent to which robots or devices make customers feel that they are in the company of another social entity and can work in conjunction with or substitute for service employees (E.g. Human employees). Following the typology of technology infusions into customers’ service frontline experience, two types of social presence can be distinguished, namely: human social presence and automated social presence. ‘Automated’ social presence emphasizes that human agents are replaced by technology as social agents. Additionally, agents are seen as human in prior research, but according to a more recent study, it can also be a service agent in a technological sense (I.e. service robots) (Van Doorn et al., 2017). That is important to note, because there is little research done on this phenomenon (Sabelli & Kanda 2015). However, service robots cannot be physically excluded from, but have to cooperate with humans (Rechtsteiner, Thaler & Troester, 1995; Iborra et al. 2009). Regarding to the usage of service robots, different appearances of robots can be distinguished (Walters et al. 2008). As suggested, robots can be placed on an anthropomorphic appearance scale which varies from very human-like to very machine-like appearance (Goetz, 2003; Walters, 2008). An explanation of the most separate robots in appearance.

Human-like-robot: Human-like robots will possess certain human-like features, these features are mostly stylized or simplified version of the human equivalent. The robots have the basic appearance of the human basic anatomy, which means; having a face, having arms with hands and fingers, and have a standing position on legs (Miklósi & Gácsi, 2012; Gong & Nass, 2007). Besides, it may have eyes, ears, eyebrows and wheels (Walters et al. 2009).

Machine-like-robot: A machine-like-robot looks different compared to the human-like robot. It is relatively machine-like in appearance and it will not have human-like features (Walters et al. 2009; Gong & Nass, 2007). Additionally, it can communicate by making movements with the body of the robot (Nicolescu & Mataric, 2001).

research which investigated the habituation of people and robots. Their results imply that people indeed get used to robots and perceptions differ over time. Additionally, there is additional evidence that humans’ perceptions about a robot is affected by the level of perceived intelligence of the robot (Eimler & Kramer, 2010; Bartneck et al. 2009). In general, higher level of intelligence and competence is ascribed to more human-like robots. Accordingly, machine-like robots are lower rated on utility, competence, and intelligence in general than human-machine-like robots (Canning, Donahue, & Scheutz, 2014; Walters et al. 2008). Moreover, there is evidence which shows that trust is driven by the appearance and physical characteristics of the robot (Adams et al. 2003). In general, human-like appearance is inspiring more trust compared to machine-like appearance (Adams et al. 2003; Li, Rau & Li, 2010; Broadbent et al. 2013). When diving into the literature of health care and robots, there is contradictory evidence for the robots’ appearances. Depending on the care task, a certain type of appearance is preferred. In general, younger people prefer human-like robots for leisure situations in care (Scopelliti et al. 2004). However, children with autism prefer a machine-like appearance over a human-like appearance when interacting (Robins et al. 2004). Although, Dautenhahn et al. (2005) argued that appearance is less important. He claims that human-like communication is preferred when having a robot companion.

Table 1 – Literature about influence appearance in HRI

Human-robot interaction literature

Factor Conclusion of paper

Appearance

People prefer human-like robots (Walters et al. 2008)

Human-like appearance of robots is inspiring more trust (Adams et al. 2003) People preferred machine-like service robots (Khan, 1998)

Peoples preferences for a certain robot changes over time (Koay et al. 2007) Human-like robots are perceived as more trustful than machine-like

(Broadbent et al. 2013)

Human-like robots are perceived as more trustful than machine-like robots (Li, Rau, Li, 2010)

Human-like appearance is appealing to people (Hanson et al. 2005) Behavior Human-like robots are perceived as more social (Broadbent et al. 2013) Acceptation Exploration when people accept a social human-like robot (Sabelli and

Kanda, 2016)

Competences

Robots that exhibit human-like communication abilities are preferred (Dautenhahn et al. 2005)

Mechanical-like robots are lower rated on competences (Canning, Donahue, and Scheutz, 2014)

Mechanical-like robots are lower rated on abilities (Walters et al. 2008).

Care

People in different age groups give opinions regarding robots' competences (Scopelliti et al. 2004)

2.2 Human-robot interaction (HRI) research

Lately, the attention to the phenomenon HRI has increased in literature (Tsarouchi, Makris & Chryssolouris, 2016; Bartneck & Forlizzi, 2004; Bartneck et al. 2009). Tsarouchi, Makris & Chryssolouris (2016) state that HRI is the field of communication and interaction between robots and humans (Mayer et al. 2014; Moniz, 2013). Research has been focused on several HRI-topics. Brandl, Mertens, & Schlick (2016) investigated the human-robot interaction in personal service. Eimler, Krämer & Pütten (2010) conducted a theoretical framework for the interaction between 1) human and human and 2) human and robots. They argued that the interaction between human and robot is based on many of the facets that interactions between human and human embodied. That is according to Eyssel & Hegel (2012), who stated that interaction of human with technology is comparable to human-human interaction and in some cases people are not aware of the fact that they do this automatically. An important driver of how people perceive each other and interact socially with each other is by assessing personality traits of other humans (Walters et al. 2008). They argued that a similar process is observed between humans and robots. The drivers of consumer perception regarding robots are verbal and non-verbal attribute factors. One of the personality traits is the physical characteristic (Borkenau & Liebler, 1992; Walters et al. 2008). In the field of HRI there is converging evidence that physical characteristics and the robots’ appearance can have important influence on HRI (Canning, Donahue, & Scheutz, 2014; Walters et al. 2008; Adams et al. 2003; Broadbent et al. 2013; Li, Rau & Li, 2010). Thus, the appearance of the service robot is an important driver and plays a role. Butler & Agar (2001) explored the psychological effects of different robot body designs in HRI and found that appearance of the robot plays an important role. Ishiguro (2007) indicate also that consumers’ perceptions about robots might be driven by behavior and appearance. Additionally, when guiding consumers’ perceptions in HRI, the voice of the robot and the appearance of the robot as a function of social cues provides information about the robot’s persona (Powers et al., 2005; Powers & Kiesler, 2006). Concluding from the past research in HRI, especially the appearance of the robots is important.

3. Conceptual model

appearance is most important factor for customers’ perceptions when facing a service agent. The existing Academic literature does not suggest what type of appearance is preferred with regard to the service agents. Therefore, different types of robot appearances will be included in this study. According to Walters et al. (2008) service agents can be divided into two different types. The two types are a human-like and a machine-like robot and are based on the anthropomorphic scale applied by MacDorman (2006). However, van Doorn et al. (2017) suggest that it might be interesting to investigate the combination of a robot and a human in the service frontline. Hence, a human employee is the third variable in this study. There is an interesting research gap in the existing literature. According to Dawar & Parker (1994); Eyssel & Kuchenbrandt (2012) it is suggested that brand type could also play a role when evaluating robots. This study would like to investigate to what extent preferences for a certain type of agent in the service frontline is influenced by brand type. Brands can be divided into two different classes: premium and economy brands (Parment, 2008). However, when evaluating non-human agents, it is valuable to investigate the role of anthropomorphism, since robots need to rely on peoples’ tendency to anthropomorphize (Miklósi & Gácsi, 2012).

3.2 Hypotheses

3.2.1 Appearance Service agent

(Bartneck, 2009). Thus, there is a lack of familiarity when arriving at the uncanny valley (MacDorman, 2006). In literature, be (un)familiar with someone or something is called psychological distance (Hess, 2002; Edwards, Lee & Ferle, 2009). Edwards, Lee & Ferle (2009) state that familiarity and similarity are components of the psychological distance that people experience. The greater the psychological distance between people, the less familiar people are. Familiarity and similarity are used to judge another person (Edwards, Lee & Ferle, 2009). As mentioned in the theory of the uncanny valley (E.g. Mori, 1970) people become more familiar with a robot when it has the appearance of a real person. People prefer familiarity and social bonding with similar others (Baumeister & Leary, 1995; Feld, 1982). The others are judged by personal characteristics (E.g. appearance) (Feld, 1982). Moreover, people exhibit more a favorable bias towards people with similar personal characteristics and are more attracted to those people (Turner & Haslam, 2001; Mohammed & Angell, 2004).

Figure 1 – Uncanny Valley by Mori (1970)

it is also expected that people prefer a human-like robot over a machine-like robot since the human-like robot is more similar to people on the anthropomorphic scale (I.e. appearance characteristics) than a machine-like robot. Therefore, this study state the following hypotheses.

Hypothesis 1a: Consumers prefer a human employee over a human-like robot as service agent. Hypothesis 1b: Consumers prefer a human-like robot over a machine-like robot as service agent.

3.2.2 The influence of brand type

The way people see a company depends on different dimensions. These dimensions combined together are called brand equity (Ahearne, et al., 2010). Brand equity refers to consumers’ intangible and subjective assessment of an innovation (Ou, Verhoef & Wiesel, 2017). ‘’Brand equity consists of five dimensions: brand loyalty, brand awareness, perceived quality, brand associations, and other proprietary brand assets’’ (Ahearne, et al., 2010, p.131). The dimensions are the input for people to create an opinion about a company and its activities. Zinkhan & Smith (1992) define brand equity as: ‘’A set of brand assets and liabilities linked to a brand, its name and symbol, that add to or detract from the value provided by a product to that firm’s customers’’. Simon & Sullivan (1990) defined brand equity as: ‘’Brand equity is the marketing effect that accrue to a product with its brand name compared with those that would accrue if the same product did not have the brand name’’. There is an agreement among researchers that this latter is the best definition that describes brand equity (Aaker 1991, Ailawadi, 2003; Dubin, 1998; Keller, 2003). Accordingly, the latter definition will be used in this paper. Aaker (1996) states that brand name is one of the most important parts of brand equity. Brands can be divided into two different classes: premium and economy brands (Parment, 2008; Park, Milberg & Lawson, 1991). Where premium brands have a high level of brand equity, economy brands have a low level of brand equity (Aaker, 1996). The key driver of brand equity is brand image, which refers to customers’ perception and feeling about a brand and has influence on consumer behavior (Zhang, 2015; Faircloth, Capella & Alford, 2001). Brand image is significantly and positively influenced by innovations (Henard & Dacin, 2010; Hanaysha, 2016).

in the role of a service agent. Brexendorf, Bayus & Keller (2015) claims brand to be the most important factor when building a reputation for the introduction of a technological innovation. This is an important note since it implies that people think differently about a brands reputation with regard to their innovations. However, prior research about perceptions of innovations has tended to focus on products instead of services (Littler & Melanthiou, 2006). Furthermore, the influence of brand type on preferred service agent will be investigated by differentiate two types of brands; premium brands and economy brands (Park, Milberg & Lawson, 1991; Verhoef, Langerak & Donkers, 2007; Wilke & Zaichkowsky, 1999). Premium brands are ranked with higher quality compared to economy brands (Verhoef, Langerak & Donkers, 2007). Moreover, Park, Milberg & Lawson (1991) suggest that people ranked premium brands higher on expressions like status, wealth, and luxury, and it can be stated that people who like premium brand also like these expressions. Premium brands are understood as customers’ expression of self-concepts or image (Park, Milberg & Lawson, 1991). It is important that these expressions ‘’fits’’ with a brand image. Research referred to the best ‘’fit’’ when people perceive a fit between the brand image and their own personality or social identity (Rust et al. 2000). One of the characteristics of the own personality is the tendency to be innovative (Keller & Holland, 1978).

Hypothesis 2a: Premium brands amplify consumers’ preference for a human-like robot over a machine-like robot.

Hypothesis 2b: Economy brands amplify consumers’ preference for a machine-like robot over a human-like robot.

3.2.3 The tendency to anthropomorphize

Anthropomorphism is the process whereby people imbue the imagined or real behavior of another entity with motivations, underlying mental states, humanlike features, motivations, or intentions (Epley et al. 2007). The past literature in the area of anthropomorphism can be seen from different perspectives. The first perspective is about the shape and the appearance of the targets. The shape of an entity is important, when it is more human-shaped, there is a higher change of anthropomorphizing (Bartneck, 2009; Graham & Poulin-Dubois, 1999; Kiesler & Goetz, 2002). Additionally, movement of an object gives the impression that a product is alive (Tremoulet and Feldman, 2000). There are more features that could have a positive influence on anthropomorphizing, like; voices, communication ability, and facial features (Dennett, 1996; Aggarwall & McGill, 2007). Additional research states that the essence of anthropomorphism is the way of perceiving humanlike characteristics in real or imagined nonhuman agents. These nonhuman agents can be everything that acts with, for example, technological gadgets or mechanical devices, robots or computers (Epley et al. 2008; Bartneck, 2009). Anthropomorphism is directly relevant to all human-computer interaction, which encompasses artificial intelligence, engineering, computer science, and various technologies (Waytz, Cacioppo & Epley 2010; Epley, Waytz, & Cacioppo, 2007).

identifying who is more likely to treat nonhuman-agents or other entities as humanlike (Waytz, Cacioppo & Epley, 2010). Waytz, Cacioppo & Epley (2010) designed the Individual Differences in Anthropomorphism Questionnaire (IDAQ) and demonstrates that differences in consumers’ propensity to anthropomorphize results in differences in behavior and how people judge each other. Moreover, they argued that people with a higher level of tendency to anthropomorphize nonhuman agents, are more likely to trust technology more with tasks.

Since the relationship between the appearance of a robot and a preferred service robot is stated in the first hypothesis, the third hypothesis can be formulated regarding the influence of consumers’ propensity to anthropomorphize. Waytz, Cacioppo & Epley (2010) argues that the higher the consumers’ propensity to anthropomorphize, the more a person will trust a robot with tasks. Robots need to rely on peoples’ tendency to anthropomorphize (Miklósi & Gácsi, 2012). Therefore, this study expects that a higher level of tendency to anthropomorphize will result in a higher preference of a robot. Moreover, it is expected that the higher the tendency, people are more likely to see human-like aspects, and therefore prefer a human-like robot over a machine-like robot. This results in the following hypotheses:

Hypothesis 3a: The higher consumers’ propensity to anthropomorphize, the stronger the preference for a human-like over a machine-like robot.

Hypothesis 3b: The higher consumers’ propensity to anthropomorphize, the weaker the preference for a human over a human-like robot.

3.3 Control variables

therefore level of education will be taken into account as well. A visual framework of the abovementioned concepts is presented below in a conceptual model (Figure 2).

Figure 2 - Conceptual Model

4. METHODOLOGY

In the methodology section the following methods will be described; the methods used, the way of data collection and a plan of analysis. First, the conjoint analysis will be explained since the customers’ preferences need to be measured. Followed by the procedure of data collection and the estimation part in which the choice based conjoint analysis (CBC) will be described. Thirdly, the implementation of the attributes and associated levels are discussed together with the measurements of the moderators. Fourthly, the choice design (I.e. design of the CBC) is outlined. As a result, a final model will be presented and its functions.

4.1 Measurement for consumers’ preferences

(2010) a personal interview can obtain value for different types of research, especially when investigating the underlying causes. Nonetheless, an advantage of the CBC is that choices are an integral part of people’s daily life, this makes it an effective method (Eggers & Sattler, 2011). Therefore, the CBC method will be used for this study. Another major benefit of the CBC is the possibility to increase the realism. In addition to this, the no-choice-option should be included in the method (Eggers & Sattler, 2011). The no-choice-option is included as separate question and is known as the dual response choice design (Wlömert & Eggers, 2016). The structure of the different steps in CBC-studies is derived from Eggers & Sattler (2011). Before the CBC analysis will be conducted, a test panel (N=5) will check the study design (I.e. attributes, attribute levels). The panel may will provide adjustments and ideas to ensure that the design of the CBC is correct.

4.2 Study design

In this part, the attributes and the attribute levels are discussed. Furthermore, the measurements of the factors that influence the relationship between type of robot and preferred service are described.

4.2.1. Attribute and attribute levels

The most important step in the set-up of a Choice Based Conjoint (CBC) is the design of the study. There should be only relevant attributes included in the conjoint analysis. The relevant attributes in this study are chosen based on literature and are presented in the next section. The number of attributes that should be included in a conjoint analysis is between three and six (Eggers & Sattler, 2011). When following the number of attributes, there can be controlled for the low efficiency of choices and it will not be too complex for processing information for respondents. The study should control for the number-of-levels-effect (NoL-effect), because the decompositional approach has been selected. The NoL-effect appears when the number of levels are not distributed equally among attributes (Eggers & Sattler, 2011; Hair et al., 2009). The number of levels should be reasonable (I.e. between three and five per attribute). There has been chosen for three levels per attribute with a constant interval, since it is known that parameter estimates will become less reliable when there are too many attribute levels included (Eggers & Sattler, 2011).

are based on the anthropomorphic scale of MacDorman (2006). In line with that scale, the three levels of appearance are used by means of visual presentation.

The second attribute is brand type. In order to measure the influence of premium and economy brands, this attribute is included. A third brand is included to increase realism and to control for the NoL-effect (Eggers & Sattler, 2011). The attribute consists of three levels; Apple (premium brand), Huawei (economy brand) and Sony (middle brand). This study is estimated to test the premium brand compared to the economy brand. The ranking of the premium, economy and middle brand is based on the best global brands list of 2016 (Interbrand, 2017). Apple has the highest ranking and is number 1, whereas Huawei can be found in the lower regions at rank 72. There has been chosen for Sony as ‘’middle’’ brand, since the position on the list (E.g. Interbrand) is rank 58. The differences in ranking is only fourteen places. However, the brand value of Sony is almost 1.5 times higher compared to Huawei. The list used is based on brand management, finance and strategy (Interbrand, 2017). Moreover, the results of the test panel indicate that people see Sony as a ‘’middle’’ brand. After asking the test panel which brand they would rank between Apple and Huawei, four out of five people named the brand Sony.

Moreover, the third attribute is discount. A third variable is included in this study to increase the realism (Eggers & Sattler, 2011). This attribute is discount and is only used to test the realism. Discount consists of three levels; no discount, five percent, and ten percent. There has been chosen for discount in percentages, since it is not wise and useful to take a concrete amount of money into account. The reason for this is that a discount of ten euro’s is a relative small discount for a television or smartphone, but a large discount for a phone cover. In table 2 an overview of the attributes and its related levels is presented.

Table 2 - Attributes and levels used in Choice Based Conjoint analysis Attribute Description

attribute Attribute level 1 Attribute level 2 Attribute level 3

Type of service agent

What does the service agent looks

like? Human employee

Human-like robot Machine-like robot Store of brand

The brand for which the service agent is employed

Sony Apple Huawei

Discount

The discount a consumer perceives when

Figure 3 - The appearances of the three different types of service agents

Human-like robot Human employee Machine-like robot

In Figure 3 the three types of service agents are visual presented. The human employee, the human-like robot and the machine-like robot. The human employee is a male, since it is assumed that people expect a male employee in a store for electronical devices. The human-like robot has a head, arms and fingers. Therefore, the human-human-like robot looks more human than the machine-like robot (Miklósi & Gácsi, 2012; Walter et al., 2009).

4.2.2 Measurement propensity to anthropomorphize

anthropomorphize by identifying four classes of commonly anthropomorphized entities. After identifying, the anthropomorphized traits are paired with the four classes. The mental state attributes used in the IDAQ are captured in previously used measures of attribution of human uniqueness and higher order cognition to human targets (Haslam et al., 2005; Demoulin et al., 2004). Waytz et al. (2010) with questions about individuals’ differences in anthropomorphism is adjusted by deleting the part ‘To what extent’, since the scale is improved from a 1-10 scale, to a 7-point Likert scale (Appendix B).

4.2.3 Control variables - Measurement of control variables

Beside the gender variable, age will be taken into account as control variable. The valid age classes that will be used, are derived from Judge (1995). The control variable of level of education will be measured by a valid scale of Snibbe, Conner & Markus (2005). Consumers’ technology readiness will be measured on the hand of the valid scale used by Parasuraman, (2000); Parasuraman & Colby (2015). This scale contains sixteen items about the relation between people and technology. All measurements of the control variables can be found in Appendix A.

4.3 Choice design

This study will contain the design criteria stated by Eggers & Sattler (2011). The choice design has a direct impact on the reliability of the study. The most difficult criterion is the utility balance. There should be no dominant alternative present in the choice set. A rising phenomenon is the Polyhedral adaptive conjoint estimation method to control for too dominant alternatives, but there is no available algorithm that is efficient enough (Eggers & Sattler, 2011). Furthermore, this study will make use of a fractional factorial design. It will become too complex for respondents to answer all possible stimuli (33). According to Eggers & Sattler (2011) the number of stimuli shown to a respondent should be between twelve and fifteen.

Another rising phenomenon in literature and an advantage of the choice based conjoint is Craft. Craft means that realistic images will be used in conjoint analysis. Craft can be divided into lower craft and higher craft (Eggers, Hauser & Selove, 2016). Higher craft means pictures or images with good quality. In this research, there has been chosen for higher craft without an incentive alignment. An incentive is not possible for product innovations, since the alternatives in the set are not available on the market yet (Eggers & Sattler, 2011).

Participants were told first that they will be exposed to different choice sets of stimuli. Additionally, the respondents are provided with extra information about the attributes and levels. The choice set consist of two options each time it will be exposed. Participants make a trade-off among the sets (Hair et al., 2009). In most cases, the CBC contains three choices in each set. But to control for the dominating role of the premium brand (i.e. Apple), the number of choices per set is reduced to two options. At two choice-options, it could be the case that people are not exposed to the dominant brand. Besides, the choice sets are random allocated and in order to create realism there is a real-life scenario created.

4.4 Data collection

The data collection is done over a period of three weeks. The respondents that are targeted for the survey and choice-based conjoint analysis (CBC) live in Europe. The largest part of these respondents lives in the Netherlands. The questionnaire is tested on a small sample (N=5) before it was sent to other respondents to ensure that all participants understood the questions about the different types of service agents (E.g. appearance of the robots), brands, anthropomorphism and the questions about technology readiness. According to Simmons & Esser (2000) the respondents were asked to think about the tenor of the questionnaire. The questions the respondents had to answer included topics like, whether the questions in the survey were stated clearly and whether the given information was sufficient or not. Because it was not always clear for the panel what to do in the survey when reading the introduction, the structure, the introduction, and the scenario were adjusted. Moreover, the structure was not always logical. Therefore, adjustments have been made to overcome these problems. Additionally, a snow-ball sample was used for this study. There is chosen for platforms like Twitter and Facebook since these platforms have a high opportunity to provoke a snow-ball effect (Peever et al., 2012). Besides, people were requested to share the questionnaire in their own network and a hyperlink to the survey was placed on the time-line of Facebook and Twitter during two weeks by multiple persons.

4.4.1 Procedure

of the attribute level (E.g. service agents, brands, and discount). After that the following scenario is presented.

Scenario.

‘’ Imagine that you want to buy a new smartphone. You have the possibility to choose

between different stores in the city center where you can buy the new smartphone. In addition, you have to make a choice between the type of employee. The employees or so-called 'service agents' in these stores are different in appearance.’’

On the page where participants distinguish between different choice sets, the following scenario will be showed (Figure 4). It becomes clear for the participants that the service agents are the same, but only differ in appearance.

5. RESULTS

In the first part of this chapter, the data will be described and prepared by making some basic analyses and plots to capture outliers, missing values, and oddities. In the second part the model fit will be discussed and finally the main effects are presented.

5.1 Data preparation

Hair et al. (2009) state when using a conjoint analysis, a sample size of fifty respondents can provide insights about consumers’ preferences. However, normally it can have serious consequences when the sample size is small (N<50). A more ideal scenario is a sample size of approximately 150 respondents. The number of respondents in this research (N=157) is sufficient to state conclusions about consumers’ preferences. Since, the data set is cleaned by deleting the cases obtained from respondents that did not fill in the right answer for the validity question (E.g. In order to check whether you are paying attention, fill in 6] Agree) (N=6). The number of cases that will be used for analyses is clean and valid (N=146).

5.2 Description sample

Before starting with the outcomes of the analysis, a sample description is presented. As mentioned in chapter 4.1, the final number of participants that passed the validation check (N=146) will be used for the results. In this research, 80 males (54,8%) and 66 females (45,2%) participated. Most of the respondents have a background of a Bachelor degree or higher (82,9%). Most people are from the Netherlands (89%), and Germany (3,4%). When looking at the age categories, it becomes clear that almost all participants are younger than 35 years old (90,4%). Thus, from the characteristics it is known that the participants are highly educated and are relatively young. Therefore, it could be the case that lots of students participated in this study.

5.3 The Utility, Logit & MNL model

unobservable latent constructs (McFadden, 1974). These constructs are represented by (V) which is the systematic component and by (ε), which is the error term. The error term captures all factors which influences the model and for which the model does not account for. The model (1) is represented as:

U = V + ε

In the choice-based conjoint vector (V) is the parameter that links the attributes of services or products (X) to preference estimates (β). The consumers’ preferences in the model are estimated by using the multinomial logit model (MNL) (Eggers & Eggers, 2011), which translates consumers’ preferences into probabilities (P). The error term is assumed to be independently distributed. The likelihood that an alternative (a) has been chosen out of (J) alternatives is presented in model (2)

𝑝 𝑎|𝐽 = exp(𝛽𝑘∙ 𝑋𝛼) exp(𝛽𝑘∙ 𝑋𝑗)

1 123

The Aggregate Logit model is presented below. This model represents the utility (U) of all respondents of the preferred service agent in the frontline (j). That means the total sum of the parameter (β) of the different products attributes across the attribute levels (k) is presented. The Logit model (3) is presented as follows

𝑈𝑗= βk Xkj+ εj

<

<23

As mentioned earlier, the Latent Class analysis will be applied in this study. The Latent Class procedure cover the heterogeneity of respondents since it allows consumers’ individual-level estimates (Jain, Bass & Chen, 1990). The analysis represents the probability that a respondent (i) will choose an alternative from the choice set J. Therefore, consumers’ individual choice probabilities can be expressed as Pi in model (4)

𝑝𝑖 𝑎|𝐽 = exp(𝛽𝑖𝑘∙ 𝑋𝛼) exp(𝛽𝑖𝑘∙ 𝑋𝑗)

1 123

5.4 Dimension reduction and reliability test

common used technique to reduce the dimension of a factor is the PCA (Malhotra, 2010). In this research, the items used for measuring the control variables are originating from prior research. However, the scale of technology readiness is multifaceted, comprising four dimensions (Optimism, Innovativeness, Discomfort, Insecurity). These dimensions together are called ‘’technology readiness’’ (Parasuraman & Colby, 2015). Furthermore, to test the factors on reliability a reliability test will be performed. Aim of estimating a reliability test is to find out what percentage of the variability in the scores is due to measurement errors and what percentage is due to variability in true scores. This analysis will be performed for the dimensions of the ‘’propensity to anthropomorphize’’ and the ‘’Technology Readiness’’. Firstly, a reliability test for the moderator ‘’Propensity to anthropomorphize’’ is conducted. The Cronbach’s Alpha (CA) is good (>.6). Only the item ‘’The average fish has a free will’’ could be deleted, since the CA should increase from 0,865 to 0,875. However, the CA is already high, therefore the item is not deleted (Appendix C). Secondly, compared to the first moderator, Technology Readiness does not meet all the criteria of the PCA. The CA for this variable is .608 with sixteen items. According to Field (2009), this score is mediocre. When looking at the outcomes, it becomes clear that all items of ‘’Insecurity’’ could be deleted to increase the CA. However, this study will not exclude the whole factor because the items are important for this study. Hence, only two items are excluded to increase the Cronbach’s Alpha. The items ‘’Technology lowers the quality of relationships by reducing personal interaction’’ and ‘’I do not feel confident doing business with a place that can only be reached online’’ are excluded. Consequently, the score of the CA increased to .665 on 14 items. Moreover, after conducting a PCA, the items load on the same four dimensions. The cumulative percentages are sufficient (67,89%) and only eigenvalues higher than one are taken into account, which results in four components (Appendix C) (E.g. Parasuraman & Colby, 2015). Besides, the KMO (.749) and Bartlett’s test (.000) are sufficient (Malhotra, 2010). Since the results of the factor analysis suggest that the items load on four different dimensions, a regression analysis to test for multicollinearity is performed. Accordingly, the low VIF-scores (<1.5) indicated that there are no problems with respect to multicollinearity. The four components are included as control variables in the second and third model (Paragraph 5.5.2).

5.5 Aggregate Choice-based Conjoint Analysis 5.5.1 Type of model

For discount the variable could be interpreted as a linear or part-worth model, so this needs to be further investigated. Based on the outcomes (table 3), it is better to estimate the variable discount as a linear model. The Likelihood ratio test shows that the linear and part-worth model do not differ significantly (P>.10). Therefore, the linear model (with less parameters) will be used for discount. Furthermore, the scores of the McFadden R2 are quite similar in both models and the goodness of fit is acceptable. Accordingly, the model accounts for individual differences (R2adj.>.2). The outcomes of both models are presented in (table 3).

Table 3 - Comparison part-worth and linear model for discount. *McFadden’s R2

Attribute LL(0) LL(ß*) Number of parameters Degrees of Freedom Likelihood Ratio test PseudoR2 PseudoR2 adj. Discount (linear) -2833,59 -2275,04 8 _{138 p(Chisq=1,32;} Df=2)>0.10 0.197 0.200 Discount (Part-Worth) -2833,59 -2274,38 10 136 0.197 0.201 5.5.2 Model fit

Table 4 – Model fit

Model 1- Main effects only *Model 2 - Main effect and moderators **Model 3 - Main effects and significant moderators LL(0) -2833,59 -2833,59 -2833,59 LL(β*) -2275,04 -2265,91 -2268,91 Parameters 6 13 9 Likelihood Ratio Test compared to model (number -1) - _{P(Chi2=18.26;} Df=7)<.05 P(Chi2=6.00; Df=4)>.05 Likelihood Ratio Test

compared to LL(0) P(Chi2=1117.1; Df=140)<.0001 P(Chi2=1135.36; Df=133)<.0001 P(Chi2=1129.36; Df=137)<.0001

McFadden R2 0.197 0.200 0.199 McFadden Adj. R2 0.199 0.205 0.203 BIC 4579,9758 4596,6067 4582,6751 AIC 4562,0742 4557,8198 4555,8226 AIC3 4568,0742 4570,8198 4564,8226 CAIC 4585,9758 4609,6067 4591,6751 Hit rate 71,16% 71,69% 71,48%

* Likelihood Ratio Test compared to model 1. **Likelihood Ratio Test compared to model 2.

5.5.3 Main effects

The main effects from model 3 are given in table 5. In Appendix D, the full results of the main effects are presented. In the table, the scores of the utility, z-values, Wald-statistics, p-values, and the relative importance of the attributes are given. The Wald statistics gives the information that all attributes are significant at an Alpha level of 1%. The most important attribute is the brand type (51,71%). Discount has a relative importance level of 25,67% and the least important attribute is the appearance of the service agent (22,62%). Accordingly, participants do not always like the combinations in the choice-options since the none-option is significantly positive (ß=.5089; p<.001). This means that people do not always choose the preferred option.

The first aspect that was investigated is the appearance of the service agent. In hypothesis 1a, it is argued that people will prefer a human employee over a human-like robot as a service agent. The results of the t-test indicate that the differences between the utilities of a human employee (ß=.2848) and the human-like robot (ß=-.2378) is significant (t=6,57; p<.001). That means that hypothesis 1a is supported. In hypothesis 1b, it is argued that people will prefer a human-like robot over a machine-like robot. The results of the t-test indicate that the differences between the utilities of a human-like robot (ß=-.2378) and the machine-like robot (ß=-.0470) is indeed significant (t=2.40; p<.05). Nevertheless, hypothesis 1b is not supported since a machine-like robot is preferred over a human-like robot by consumers.

Table 5 – Main effects model 3

Model 3 (Main Effects)

Attributes Utility Wald p-value

Appearance Relative importance = 22,62% Human Employee 0,2848 49,0373 p<.001 Human-like robot -0,2378 Machine-like robot -0,0470 Brand Relative importance = 51,71% Premium 0,6634 322,9921 p<.001 Economy -0,5315 Middle -0,1319 Discount Relative importance = 25,67% 0,5932 312,9921 p<.001 None_option 0,5089 32,5609 p<.001 5.5.4 Moderating effects

a machine-like robot, the study suggests that economy brands do have a negative influence on consumers’ preferences for the human employee 0,1955; p<.05) and human-like robot (ß=-0,1338; p<.05). Hence, hypothesis 2a and hypothesis 2b are not supported.

The second moderating effect that was analyzed is whether the tendency to anthropomorphize affect the relationship between preferred service agent and appearance of service agent. In hypothesis 3a, it is stated that the higher consumers’ propensity to anthropomorphize, the stronger the preference for a human-like over a machine-like robot. Unfortunately, it cannot be said that the people who have a higher level of tendency to anthropomorphize do prefer a human-like robot over a machine-like robot (p>.10). In hypothesis 3b, it is argued that the higher consumers’ propensity to anthropomorphize, the weaker the preference for a human employee over a human-like robot. The people who have a higher propensity to anthropomorphize have the tendency to prefer the human employee (ß=.0582; p<.10). Accordingly, hypothesis 3a is not supported, where hypothesis 3b is supported.

Table 6 – Moderating effects

Model 2 (Main effects + moderators) Model 3 (Main Effects + significant moderators)

Attributes Utility Wald p-value Utility Wald p-value

Anthropomorphism x human-employee 0,0875 5,5445 P<.05 0,0582 3,3979 P<.10 Anthropomorphism x human-like robot -0,0557 2,2747 P>.10 *Anthropomorphism x machine-like robot 0,0318 Premium x human employee 0,1165 3,0943 P<.10 0,1177 3,1605 P<.10 Premium x human-like robot -0,1215 3,6272 P<.10 -0,1255 3,8812 P<0.5 *Premium x machine-like robot 0,050 Economy x human employee -0,1961 9,2798 P<.05 -0,1955 9,2362 P<.05 Economy x human-like robot -0,1318 3,937 P<.05 -0,1338 4,0594 P<.05 *Economy x machine-like robot 0,3279

* is the reference level, there are no p-values conducted for this item and therefore is not significant.

5.5.5 Predictive validity

aggregate model (McCullough, 2002). Measuring the predictive validity of model 3 should be done by making use of the holdout sample. These holdout choices are excluded when estimating the results and are only used for the predictive sample. MAE’s below five points will reflect acceptable good fitting model (McCullough, 2002). The absolute error of every alternative is calculated by the difference between the observed and predicted share. In table 7 the outcomes of the MAE are presented. For this research, the mean absolute error is 7,34%. This is above the five points, but still reasonable since the number of options is two in this case instead of more than two options. In a case with two options it is harder to predict shares compared to a case with multiple options.

Table 7 – Mean Absolute Error

1 2 Total

Observed shares 42,47% 57,53% 100% Predicted shares 35,13% 64,87% 100% Absolute Error 7,34% 7,34% 7,34%

Mean Absolute Error (MAE) (7,34+7,34/2)=7,34% 5.5.6 Control variables

The control variable level of education is divided into two groups. The first group of people with a minimum of Bachelor Degree (N=121) and the other group consist of people with a professional education degree or lower (N=25). This is executed in order to control for a decrease in Degrees of Freedom. Instead of five groups, the variable level of education contains only two groups. This means a save in degrees of freedom, since the number of parameters has decreased. Furthermore, in the Latent Class analysis (Chapter 5.6.1), the control variables will be taken into account. Moreover, age, gender, and the four dimensions of technology readiness will be used in order to explain consumer heterogeneity.

5.6 Preference-based segmentation

A drawback from an aggregate logit model is that there is no preference heterogeneity across consumers. A second disadvantages is that the logit model assumes independence of Irrelevant Alternatives (IIA). Thus, the prior presented choice model neglect heterogeneity, which means that all consumers are clones. Therefore, a Latent Class analysis is conducted, since this captured consumer heterogeneity.

5.6.1. Latent Class analysis

in this study Latent Class is better since consumers’ preferences are measured which are latent. Besides, the Latent Class analysis is able to address both of the two drawbacks of aggregate logit. There are two ways of estimating procedures that account for heterogeneity; priori segmentation, and mixture models. The mixture models are applied in this study, since it is not known which socio-demographic information should be used for segmentation (priori method).

The idea of latent class is that every consumer is allocated with a certain probability to a segment. The model will be conducted with only significant moderators and control variables as covariates. The final model will be used in this study since the outcomes are better (E.g. classification error, information criteria). First, it becomes clear that the classification errors are all, except from five classes, below five percent (Table 8). A score below five percent means that the model allocates respondents very well. In this final table only the significant moderators and control variables are included. Secondly, from table 8 it can be extracted that the scores of BIC, AIC, AIC3, and CAIC are lower when adding the only significant moderators and control variables. This is good, since the score should be low as possible. In particular, the focus is on the lowest score of BIC and CAIC, because those are preferred for larger sample sizes and provide higher penalties for complexity (i.e. Latent classes). According to the information criteria (i.e. BIC, CAIC), six segments will be used (CAIC=4181,34; BIC=4122,34). The classification error is 4,16%. Thus, there is 4,16% probability that respondents will not belong to the most probable segment. Accordingly, in this case it is a good score of the classification error. Thereby, the number of Degrees of Freedom are still sufficient. Lastly, the scree plot (Appendix D) is observed. It shows a tipping point for CAIC and BIC at six classes. It can be concluded that six segments are indeed the optimal number of segments.

Table 8 - Segments with only significant moderators and control variables as covariates

BIC(LL) AIC(LL) AIC3(LL) CAIC(LL) Npar df Class.Err.

2-Class Choice 4410,2863 4365,5322 4380,5322 4425,2863 15 131 0,0235 3-Class Choice 4274,7759 4197,2021 4223,2021 4300,7759 26 120 0,0438 4-Class Choice 4193,5457 4083,1523 4120,1523 4230,5457 37 109 0,0421 5-Class Choice 4137,5846 3994,3715 4042,3715 4185,5846 48 98 0,0523 6-Class Choice 4122,3445 3946,3117 4005,3117 4181,3445 59 87 0,0416 7-Class Choice 4121,3329 3912,4804 3982,4804 4191,3329 70 76 0,0434 8-Class Choice 4129,5772 3887,905 3968,905 4210,5772 81 65 0,0457 5.6.2 Interpretation of segments

negative, it means that consumers are not willing to choose the no-choice option. Since the parameter discount is linear, it is important to find a common scale. The relative importance is calculated after finding common scale. Note: The segment that accommodate most people is not always the most valuable segment. Furthermore, level of education and age are not significant at all.

Segment 1 (Main stream technology based women)

This is the largest segment. The number of consumers in this segment is 47. The score for insecurity is relative high compared to segment two. Important note is that the most people in this segment are female and have a relatively high level of tendency to anthropomorphize. People in this segment are more aware of the appearance, since they attach more value to it (11,58%). They prefer the human-employee (ß=.8158) over the other service agents. The second preferred type of agent is the machine-like robot (ß=-.3778). The premium brand is most preferred (ß=1.9447). The brand type is relatively important for this group (31,44%). Nevertheless, this group did not choose the preferred option since the sign of the no-choice option is positive (ß=2.1638). Hence, people in this segment prefer a human employee working for a premium brand.

Segment 2 (Brand lovers with high technology readiness)

The number of consumers in this segment is 31, which means it is the second largest segment. The score of insecurity is relatively high compared to segment 3,4 and 5. To summarize this segment, the most people for this segment are male. People in this segment consider brand type as an important factor (35,14%). They prefer the premium brand (ß=1.6038) over the economy brand. The human-employee is still preferred, but they do not attach any value to the appearance of the agent (2,85%). However, people in this segment still want a human-employee from a premium brand. To conclude, these people do know what they want since the no-choice option is not preferred (ß=-2.1123).

Segment 3 (Cheap dogs)

and the level of insecurity of people is relative low. These consumers want a machine-like robot from the economy brand.

Segment 4 (Bargain Hunters)

The number of consumers in this segment is 19. These people do like the preferred option (ß=-4.8678). They prefer still the human employee, but important is that these people dislike the human-like robot the most. The relative importance of the appearance is low (5,47%). Interesting is the fact that they prefer the economy brand (E.g. Huawei). They do attach more value to brand type than they do attach value to appearance. Additionally, this segment contains people which are aware of price (ß=2.1766). Moreover, the relative importance of discount is high (85,23%). Furthermore, like most segments there are slightly more male than female. The level of insecurity is relative low compared to the other segments. Thus, people prefer a human employee from the economy brand.

Segment 5 (Keep the status-quo)

The number of consumers for this segment is 18. This is the only segment were people consider appearance as really important. These consumers prefer a human-employee over the other types of services agents. They prefer a human employee (ß=2.0594) over the other two service agents and they do not like the machine-like robot (ß=-1,1749). This is important to know, since the relative importance is bigger for type of service agent (44,62%) compared to the type of brand (20,47%) and discount (34,92%). The economy brand is preferred more compared to the premium brand. Besides appearance, the discount is considered an equally important for this segment. There are slightly more females in this segment and have a high tendency to anthropomorphize. Concluding, these people want a human employee as service agent from an economy brand.

Segment 6 (Low educated for who service agent doesn’t matter)

The level of insecurity is negligible since the scores are both high and low. Hence, people in this segment prefer a machine-like robot as service agent from the premium brand.

Conclusion is that according to this study, the most important segments are segment 1 and segment 5. These people consider appearance of the service agent equally important. Furthermore, it seems that the segments who prefer a human-like or machine-like robot over the human employee, are not interested in appearance at all. This is a result of the low scores for the relative importance (table 10).

Table 9 – Preferences per segment

Attributes Class1 Class2 Class3 Class4 Class5 Class6

Appearance Human employee 0,8158 0,1509 -0,0863 0,8358 2,0594 0,0989 Human-like robot -0,438 -0,0897 -0,2168 -0,5364 -0,8845 -0,2979 Machine-like robot -0,3778 -0,0612 0,3031 -0,2993 -1,1749 0,1990 Brand Premium 1,9447 1,6038 -0,3046 -0,2223 -0,7174 0,9329 Economy -1,4584 -1,3634 0,2188 0,2454 -0,0488 -0,8186 Middle -0,4863 -0,2405 0,0858 -0,0231 0,7663 -0,1143 Discount 0,6166 0,5235 1,1706 2,1766 0,2531 1,7095 None_option 2,1638 -2,1123 2,0471 -4,8678 -0,365 6,3573

Range Class1 Class2 Class3 Class4 Class5 Class6

Appearance 1,2538 0,2406 0,5199 1,3722 3,2343 0,4969

Brand 3,4031 2,9672 1,4752 2,3989 1,4837 2,5281

Discount -6,166 -5,235 -11,706 -21,766 -2,531 -17,095

Totaal 10,8229 8,4428 13,7011 25,5371 7,249 20,12

Table 10 – Relative importance per segment

Relative importance Class1 Class2 Class3 Class4 Class5 Class6

Appearance 11,58% 2,85% 3,79% 5,37% 44,62% 2,47%

Brand 31,44% 35,14% 10,77% 9,39% 20,47% 12,57%

Discount 56,97% 62,01% 85,44% 85,23% 34,92% 84,97%

5.7 Overview of tested hypotheses

Table 11 – Hypothesis supported?

Hypothesis Supported?

1a Consumers prefer a human employee over a human-like robot as service _agent. Yes 1b Consumers prefer a human-like robot over a machine-like robot as service

agent. No

2a Premium brands amplify consumers’ preference for a human-like robot over a

machine-like robot. No

3a The higher consumers’ propensity to anthropomorphize, the stronger the _{preference for a human-like over’ a machine-like robot’.} No 3b The higher consumers’ propensity to anthropomorphize, the weaker the _{preference for a human over a human-like robot.} No

6. DISCUSSION, FURTHER RESEARCH AND LIMITATIONS,

MANAGERIAL IMPLICATIONS

6.1 Discussion

6.1.1 The preferred service agent