Measuring consumer preferences for automated agents in a hotel context

(1)

Measuring consumer preferences for

automated agents in a hotel context

Differences between hedonic and utilitarian travel motives

By

Shu Han, Chuang

(2)

1

Measuring consumer preferences for

automated agents in a hotel context

Differences between hedonic and utilitarian travel motives

Master Thesis

MSc Marketing Intelligence

University of Groningen, Faculty of Economics and Business

1

st

Supervisor: Dr. Jenny van Doorn

2

nd

supervisor: Dr. Lara Lobschat

completion date: 26

th

June, 2017

Author: Shu Han, Chuang

Student number: 3077918

Address: Duindoornstraat 45, 9741NN, Groningen

Phone number: +31 63392 4760

(3)

2

Abstract

This research aims to analyze the effect of physical embodiment, the extent of anthropomorphism of automated agents and the impact of social interaction on the preference of the consumer in two scenarios (i.e. hedonic and utilitarian) in the hotel context. Based on the Choice-Based Conjoint analysis, the physical embodied robot is more preferred over the virtual agent. However, the anthropomorphism of the robot does not lead to higher evaluation. Furthermore, the passive social interaction of the intelligent agent as service encounter is more preferred over the active social interaction. The finding is contradicted to the active social interaction as expected from human staffs. Thirdly, the positive effect of anthropomorphism is stronger when consumers are driven by hedonic motives rather than utilitarian ones, relative to the physical embodied robot without anthropomorphism. Finally, latent-class analysis is conducted for segmentation to accommodate the heterogeneity of consumer preference.

Keywords: automated agent, physical embodiment, anthropomorphism, social interaction,

(4)

3

Preface and acknowledgement

This master thesis is my last destination of my master journey at the university of Groningen, MSc Marketing Intelligence. During the journey, I had some doubts concerning whether I would be able to make it to the end. Finally, I stand on top of the mountain, appreciating the beautiful scenery on top, and having the courage to embrace the new challenges in the next stage of my life. Without the support from some important people, I could not survive until now.

Firstly, I would like to thank my supervisor, Jenny van Doorn, having the patience to guide me. Since this is the first academic work I have ever had, she supported me on building the convincing train of thought and finding line of reasoning, which I was short of. Although the training process is not always pleasant, I am extremely thankful for the guidelines she had provided on the way. Secondly, I would like to thank the fellow students in the thesis group, who provided constructive advices during every meeting. Specifically, I would like to thank Hang He, who had intensive discussions with me during the entire process. Fourthly, I would like to especially thank Mark Pijper, who devoted to proof reading this thesis and checking mistakes for me besides the busy schedule. Fifthly, I want to thank my family. Whenever I have difficulty, my parents, sister and brother are the ones who encourage me the most. Last but not least, I am really grateful to 219 respondents and friends who devoted their time and energy to filling out the survey. Without every one of you, I would not successfully finish the thesis.

I hope you enjoy reading my thesis!

Shu Han, Chuang

(5)

4

1. Introduction

Automation in hotel industry

The infusion of technology is dramatically changing the customer’s service experience and the way service is delivered by service encounters. For instance, in the hotel industry, Hilton Hotels recently launched the robot, Connie, which is powered by Watson artificial intelligence from IBM, to welcome guests in Washington DC (Solomon, 2017). Instead of traditional interaction in which customers would ask any information from human staffs, the robot allows the international travelers to communicate in multiple languages. A hotel in Japan, Henn-Na, is the first hotel that entirely staffed by robots. Customers are firstly welcomed by three robots with human, human-like, dinosaur appearance at reception with check-in and info desks service. In their room, a robot in the shape of pink tulip addresses simple questions such as, “what time is it?” and “What is the weather tomorrow?” Customers can also ask the robot to turn off the lights (Guardian, 2017). Most recently, a list of corporate hotel chains (e.g. Wynn Las Vegas, Marriott International, Hyatt Hotels) has added smart speakers (e.g. Amazon Echo, Apple’s Homepod) to the travel experience (Johnson, 2017). With the intelligent assistant in the room, customers are able to order their car or request fresh towels. In the past, self-service technologies in the hospitality industry (e.g. self-check in) were unable to engage customers in the social level. Nowadays, artificial intelligence is designed with the notion of ‘automated social presence (ASP),’ i.e. the extent to which machines enable customers to perceive that they are in the company of another social entity gives more possibilities for the use of service agents in service context (Van Doorn et al., 2017).Consistent with the idea that robots have become increasingly popular in service industry, IFR World Robotics (2016) forecasts that the estimated value of sales for professional service robots will amount to US$ 23.1 billion between 2016 and 2019, relative to US$ 4.6 billion in 2015. The forecast indicates an increasing demand from 2016 to 2019. More and more hotels adopt new technology mainly to effectively customize service offerings and spontaneously delight customers while saving labor costs simultaneously (Bitner et al., 2000). Therefore, it is important to understand preference of the consumer on new technology (e.g. robot or intelligent assistant) and meet their expectations accordingly.

Virtual agent

(8)

7

superior user-centric systems, ‘intelligent assistants’, is needed. Research has pointed out that intelligent assistants (e.g. Siri) are capable of having natural interactions with human users and information environment (Santos et al, 2017). The demonstration of the invention can be seen in the medical field. The surgeons are able to interact with the operating room assistants as if they are active members of the team via natural modalities such as speech recognition and hand gestures (Guzzoni et al., 2007). Besides, the artificial intelligent agent is particularly useful in the negotiation scenario since the information regarding the preferences of the objects may be only elicited as the interaction progresses (Lin et al., 2008). The same research indicates the automated agent leads to significantly better individual utility and plays more effectively than the human counterparts. Although the intelligent assistants are not yet a common scene in the hotel, there are a variety of added-values for customer service ranging from making phone calls, managing schedules, finding nice restaurants, finding answers to general questions, searching information on websites or even chatting with intelligent assistants (Jiang et al., 2015).

Physical embodied robot

(9)

8

Anthropomorphism

In marketing fields, anthropomorphism (i.e. human’s tendency to attach humanlike features, emotions, and motivations to non-human objects (Epley et al., 2007)) is often used in branding (Aaker, 1997) or product design (Aggarwal & McGill, 2007). Kiesler & Hinds (2004) emphasize anthropomorphism is particularly essential in robot design due to their high anthropomorphizability. Previous researches indicate human-like robots increase user engagement and elicit users to behave social and physical conventions in their familiar ways (Koda & Maes, 1996; Reeves & Nass, 1996). Leveraging anthropomorphism contributes to more positive evaluation of the robot and longer interaction with it (Kiesler et al. 2008). However, Goetz et al. (2003) stress the embodiment should be designed in a way that corresponds to the robots’ tasks instead of taking anthropomorphism as a goal itself. Therefore, it is necessary to investigate whether human-like robots as a service encounter are desirable for consumers.

Human-robot interaction

(10)

9

Travel motives

Lastly, since holiday and business travelers are two major segments in hotel industry, it is practically relevant to find out whether their preference on robot design differs. Correspondent with marketing literature, consumers’ hedonic and utilitarian values have shown to influence customer preference, satisfaction and loyalty (Cronin et al., 2000). A study introduces the value conceptualization to the travel literature and suggests both hedonic and utilitarian travel values are essential elements of the relationship between consumer perceptions of travel destinations and their satisfaction (Babin & Kim, 2001). Consequently, the influence of two traveling motives on the preference of an automated agent will be taken into account.

To understand the consumer’s preference on the automated agent in the hotel context, two main research questions will be answered throughout this research:

1. To what extent does the physical embodiment and anthropomorphism of a robot have an impact on hotel service preferences?

2. To what extent does the level of social interaction and traveling motives have an impact on hotel service preference?

The study has three main contributions. First of all, it investigates the effect of physical embodiment and anthropomorphism on the preference of an automated agent in the hotel context. Furthermore, it studies the influence of social interaction on the preference of an automated agent in the hotel context. Lastly, it considers the impact of the hedonic and utilitarian travel motives on the preference of an automated agent in the hotel context.

(11)

10

2. Theoretical framework

2.1. Literature review

2.1.1. Intelligent assistants

The spoken dialogue systems have been invented for quite some time (McTear, 2002). It is not until recently, that the voice controlled intelligent assistants (e.g. Microsoft’s Cortana, Google Now, Apple’s Siri, Amazon’s Alexa and Facebook’s M) have become people’s partner in daily life. According to Guzzoni et al., (2007), an intelligent assistant is defined as “a software system that is able to observe and sense its environment including human communications, to analyze the situation by mapping input senses into a model of what tasks and events may be happening, and to understand and anticipate what actions will produce relevant and useful behavior.” With the intelligent assistants, users are able to ask for the request such as executing the tasks of searching information on the website, setting alarms, or making phone calls. Some intelligent assistants are even capable of chatting or playing games with users (Kobayashi et al., 2015). A recently conducted research by Google indicated that 55% of the teenagers in U.S. use voice search every day, 89% of teenagers and 85% of adults acknowledge the voice search will be a common scene in the near future (Google.Inc, 2015). Besides, the major benefit users recognize is that a spoken dialogue way of interaction is more natural and efficient for people to communicate than typing (Negri et al., 2014). Intelligent assistants allow the users to access information in a new way, which is different from traditional web search. The main differences lie in the autonomy (i.e. they act on behalf of the user to achieve the tasks without the requesting command or clicking bottoms from users) and intelligence (i.e. their ability to perform the task in a context-based and user-dependent manner (Dehn & Van Mulken, 2000). Therefore, in this study, the ‘intelligent assistant’, ‘intelligent agent’, ‘automated agent’, and ‘service agent’ will be used interchangeably. The survey about the spoken dialogue technology indicated when users communicate with intelligent assistants, they expect the assistants (1) to understand users’ intent, (2) to maintain the context of the dialogue (3) to express more complex information needs, and (4) to decrease the errors in speech recognition (Jiang et al., 2013; Kiseleva et al., 2016).

2.1.2. The embodiment of the automated agent

(12)

11

are able to demonstrate reactive and pro-active behavior in the environment) with distinct embodiments (Ziemke, 2003). As can be seen in figure 1, Milgram’s Reality-virtuality Continuum (Milgram et al., 1995) is applied to illustrate the nature of the embodiment. The robots and virtual agents can be seen as embodied at two extremes of the continuum. There are examples of the intelligent assistants that are embodied in different interfaces such as: mobile phones (e.g. Apple Siri), a smart environment (e.g. MavHome), speakers (e.g. Google Home) or even robots (e.g. Connie); however, there is lack of research comparing which type of interface would be the most preferred for the consumers.

Figure 1: Milgram’s Reality-Virtuality Continuum (Milgram et al., 1995)

2.1.3. Smart environment: virtual agent

Research in the field of Human-Computer Interaction (HCI) have presented the benefits of applying virtual agents in several settings such as taking an educational tutorial (Lester et al., 1997) and going on a tour (Isbister et al., 2000). When it comes to the ‘virtual’ agent, an early study has shown the ambiguity between verbal and non-verbal communication in similar human dialogues (Deutsch, 1974). To make a clear distinguish, relevant researches in virtual agents that communicate via voice and without any either 2D or 3D visual presentation will be discussed now. Virtual agents have been used in many occasions. The agent that serves as a bartender is designed to have conversations with visitors via natural language and to enhance the social atmosphere (Isbister & Hayes-Roth, 1998). A project in smart home context, MavHome, demonstrated the creation of an environment that performs as an intelligent agent that is able to perceive the condition of the home via sensors and respond to the environment through device controllers (Cook et al., 2003). Furthermore, the virtual agent is also applied in the meeting occasion, in which the helper agent supports human conversation by sensing contextual cues from the conversation, and then providing help accordingly rather than being present and active anytime during human conversation.

Mixed Reality

Reality Augmented Virtuality

Reality

(13)

12

Imagine that the virtual agent is placed in the hotel room, and the virtual agent turns on the heater before the alarms alert, because it has learned that the room needs 10 minutes to warm up to the adequate waking temperature. After the guest sounds the alarm, the bedroom light turns on automatically. The virtual agent starts interacting with the guests by informing the breakfast serving time in the hotel, organizing the personalized travel destinations for them, and replying any information that can be found online.

2.1.4. Speaker: Physical embodied robot

In addition to placing the intelligent assistant in an environment, they can also be embodied in the speaker. There are an increasing number of corporate hotel chains (e.g. Wynn Las Vegas, Marriott International, Haytt Hotels) and vacation rental companies (e.g. Hyatt Regency Lake Washington and AvantStay) that place speakers (e.g. Apple’s HomePod, Amazon Echo and Google Home) for guests to use in the hotel room (Johnson, 2017). In order to customize the voice service specifically for the hotel guests, a startup created a voice app, the Vacation Rental Concierge, which is available in the Alexa Skills Store or Google’s App Directory (Johnson, 2017). Besides answering the frequently asked questions such as Wi-Fi password and house rules, hotels or Airbnb hosts are able to customize their own questions according to the guest inputs such as the recommended restaurants, highlighted tourist attractions and many others on the Rental Virtual Concierge platform (Concierge, 2017).

The difference between the speaker and the virtual agent lies in the ‘embodiment’ of the intelligent assistant, which has been studied extensively in the field of social robotics. Fong et al. (2003) define embodiment as “that which establishes a basis for structural coupling by creating the potential for mutual perturbation between system and environment.” According to the definition, physical embodiment is not a mandatory process if the relationship between system and environment is not perturbative. In this paper, embodiment refers to the widely recognized meaning in the domain of Artificial Intelligence (AI) and Robotics, physical instantiation (i.e. bodily presence) (Ziemke, 2003).

(14)

13

positive impact on better affordance (i.e. the essential characteristics of a device that decides its way of use), therefore causing less frustration for people. For instance, a phone implies an intuitional cue for calling to communicate with others. Likewise, a physical embodied agent with the sound receiver and speaker may give a clue for proper social interaction (e.g. asking for information). In addition, Lee et al. (2006) conducted two experiments and concluded from experiment 1 that positive effects of physical embodiment on the evaluation of the agent, the judgement of public evaluation of the agent and the evaluation of the interaction with the agent can be explained by the feeling of an agent’s social presence. Furthermore, experiment 2 implied lonely populations feel higher social presence of social agents, and convey more positive responses to social agents than non-lonely population.

Many findings in elderly healthcare and children education support the stronger effect of physical embodied robots than virtual agents (see table 2). For instance, Mann et al. (2015) examined how people interacted with either a robot or a tablet computer providing healthcare instructions. Results found that participants had more positive interactions with the robot compared to the computer, consisting of increased speech, positive emotion (smiling), and participation in the relaxation exercise. Subsequent results revealed the robot was rated higher on scales of trust, enjoyment, and desire for future interaction.

Study Findings

Bartneck (2003) People have the same degree of enjoyment on playing a game with either a virtual or physical robot, but people scored higher with a physical robot present.

Wainer et al. (2006) People rated higher watchfulness and enjoyableness for a physical robot that was in the room with them than both a simulated robot on a computer and a real robot shown via teleconferencing.

Kose-Bagci et al. (2009) Children performed better when collaborating with a physical robot than with a virtual agent.

Looije et al. (2012) Children look more frequently and longer periods at physical robot than a virtual agent.

Leyzberg et al. (2012) A robot tutor delivers more effective educations than comparable web-based instruction or books with audio.

Kwak et al. (2013) Children were asked to exert electric shocks to either a physically present robot or a simulated robot on a computer screen, both of them showed colored bruises after being shocked. Children expressed significantly higher empathy on embodied robot than the computer-simulated one.

(15)

14 conditions.

Mann et al. (2015) People were more likely to follow relaxation instructions from a robot than from a computer tablet with the same software and voice.

Table 2: empirical evidences of physical embodiment

Previous research in social robotics has vividly demonstrated the comparison between physical embodied and virtual agents in education, teleconference or healthcare context. However, there is no research conducting the effect of physical embodiment in the hotel context even though the service is already available in the real world.

2.1.5. Human-like robot: anthropomorphized robot

As shortly discussed before in the introduction section, human-like robots (e.g. Connie) have been adopted by some hotels (e.g. Hilton, Henn-Na). The robots serve as service encounters that are able to provide verbal customer services, which can be seen as intelligent assistants embodied with anthropomorphic appearance. Anthropomorphism is the phenomenon that people attribute human characteristics to a non-human being (Hatano & Inagaki, 1994). Several studies show that humans are born with the ability to make anthropomorphic attributions instead of learning (Hatano & Inagaki, 1994). For example, we call our planet as ‘Mother Earth’ and often use it while discussing environmental issues (Wilson, 2006). Based on the common human nature, an anthropomorphic form has been used for designing technology (Duffy, 2003) and products (Aaker, 1997; Aggarwal & McGill, 2007). The attribution is also applied in HRI. To provide comprehensive understanding of anthropomorphism, factors that influence anthropomorphism, benefits of anthropomorphic robots and disadvantages of anthropomorphism will be concisely discussed in the following.

(16)

15

that mobilize with legs instead of wheels (Chew et al., 2010)). Di Salvo et al. (2002) suggest humans perceive a humanoid head as human-like due to the presence of particular characteristics and dimensions of the head. However, research shows that people who differ in age (Kamide et al., 2013), gender (Eyssel et al., 2012), culture (Evers et al., 2008) and motivation (Epley et al., 2007) have different preferences on appearance of the robot.

Secondly, research has indicated that facilitating human-agent interaction is the most critical benefit of anthropomorphism (Giullian et al., 2010; Feil-Seifer & Mataric, 2011; Wade et al., 2011; Fasola & Mataric, 2012). Besides, anthropomorphism enables humans to apply skills that developed from human-human interactions (Schmitz, 2011). There are many empirical evidences show the benefits of anthropomorphism in HRI. For instance, Riek et al. (2009) indicated people empathize stronger towards human-like robots than machine-like robots. Similarly, Bartneck et al. (2006) conducted an experiment that could give plus or minus points as praise and punishment for correct or wrong partner answers. Results show that anthropomorphized robot were praised more and punished less compared to machine-like robots and human. The insight was implemented to increase consumer’s willingness to care about robots (Bartneck et al., 2006). Furthermore, people comply more with a robot that matched the seriousness of the task (Goetz et al. 2003).

(17)

16

As has been observed so far, people’s preference differs depending on the context. However, there is lack of research investigating the preference on the embodiment of intelligent agents in the hotel context.

2.1.6. Social interaction

Social interaction between customers and employees in the service experience has been identified as important determinants of customer satisfaction and patronage behavior (Mittal & Lassar, 1996). The study also shows that these interactions are especially essential for a lonely population who counts on service encounters as their main source of human interaction (Forman, 1991). Nowadays, robots or artificial intelligent agents have become an alternative to enhance frontline service experience that involves higher level automated social presence compared to conventional self-service technology (Van Doorn et al., 2017).

It has been stressed that not only the appearance of a robot but also the behavior influences the human-robot interaction (Minato et al., 2004). In the field of social robot, Breazeal (2003) distinguishes four subclasses from existing applications: (1) socially evocative (2) social interface (3) socially receptive (4) sociable. All four classes are designed for motivating people to anthropomorphize the robots in order to interact with them. In this section, we focus on the function of social interaction, thus the latter two classes would be further discussed. Socially receptive robots are socially passive; they are not pro-actively engaging in satisfying people’s internal social goals. On the other hand, sociable robots are pro-actively engaging in people’s social needs, regarding not only help the people complete the tasks, but also improve their performance and learn from human (Breazeal, 2003).

There are many researches actively investigated whether people exhibit the same interaction behavior from human and robots. To answer the question, studies in Human-computer interaction and Human-robot interaction will be further elaborated.

(18)

17

reciprocity towards computers. By leveraging the rule of reciprocity, people show more helping behavior to a computer if it helped them first, instead of another computer which did not help them (Fogg and Nass, 1997). Furthermore, people are more willing to reveal their intimate information to computers that expose more about themselves in the initial contact than the other computers which did not (Moon, 2000). The founding from Human-computer interaction is found to be the case when people interact with robots. For instance, Lammer et al. (2014) demonstrated that older adults’ perceived the usability and ease of learning of the care robot as more positive if the robot requested for help and repay with a favor than if the robot did not show the above two behaviors. In other words, ‘mutual care’ between human and robot enhances their relationship, and consequently acceptance in the case of care robots.

(19)

18 2.2. Conceptual framework Social interaction  Active  Passive Hotel service preference Type of embodiment 1. Virtual agent

2. Physical embodied robot 3. Physical embodied and

anthropomorphized robot

Travel motives

 Hedonic  utilitarian

Control variables

• Demographics: Age, Gender, Income, Occupation, Nationality • Frequency of staying in hotel • Negative attitude towards robot

H1

H3

H4a, b H5a, b

Figure 3: conceptual framework

Price  €100 per night  €110 per night  €120 per night H2 Note:

1. ‘Physical embodied robot’ is presented by machine-like robot

(20)

19

To build a service agent that engages interaction with customers, the theory from human relationship is applied. Specifically, Eimler et al. (2010) proposes a theoretical framework integrating the ‘theory of need to belong’, ‘social exchange theory’ and ‘theory of mind’ as the basis for sociable companions and for human-artifact interaction. The characteristics and capabilities that are critical for human to engage in robot interaction are elaborated in this research. Based on the ‘theory of need to belong’, it is suggested to build the physical embodied robot that is as visible as possible, therefore the users would benefit from the condition of ‘propinquity’ (i.e. mere exposure effect). Moreover, human-like robots are recommended to make use of the second condition in the ‘theory of need to belong’, namely ‘similarity’, so that users could use the cues of the attractiveness, dress and speak way of robot that are similar to them. However, Eimler et al., (2010) mentioned it is necessary to provide empirical studies in certain context. Therefore, three types of embodiment with virtual agent, human-like and machine-like robot are proposed accordingly.

Furthermore, the social interaction needs to be taken into account when designing a service agent. In the hotel, staffs are expected to show their active interaction to the guests in general. In the ‘theory of need to belong’, the second condition ‘similarity’ suggests the service agent should be designed in a way that they behave similar to humans (e.g. show hospitality); Besides, the ‘reciprocal rule’ in the social exchange theory implies if the service agent initiates the interaction, users are more willing to build the relationship with it. Lastly, the ‘theory of mind’, which has been recognized as a basic prerequisite for human-human interaction, describes the human’s ability to take the perspective of the counterparts and predict what they think (Frith & Frith, 2003). To make use the insight from the ‘theory of mind’, the service agent should be designed with the capability to predict what users need and provide guidance before they asked for. However, the assumption based on theory of human-human relationship has not been tested in the human-agent scenario, therefore, it is necessary to take social interaction (i.e. active or passive) into account, in addition to type of embodiment.

(21)

20

additional service (e.g. friendliness of service personal) provided by the company, which influences consumption preference (Ostrom and Lacobucci, 1995). Furthermore, a higher price implies higher quality, which indicates a consumer is buying a technologically superior product or service. Nonetheless, the hypothesis between price and hotel preference will not be built forth on in this section, since the price normally leads to negative effect on the preference as indicated from microeconomic theory.

(22)

21

2.3. Hypotheses

To increase customer satisfaction and therefore be more preferred by the consumer, the service agent/robot is designed in a way to facilitate interaction and build the relations with human. Therefore, the ‘theory of need to belong’, ‘social exchange theory’, and ‘theory of mind’ derived from the human-human relationship are used as the basis to develop hypotheses for human-agent/robot interaction.

2.3.1. Type of embodiment

(23)

22

1995; Nahemow et al., 1975). Besides, Zajonc (1980) indicates people rate the objects who seemed similar to themselves more likable and more familiar. According to the studies in interpersonal relationships, it can be assumed that a robot/agent should be designed as visible and proximate as possible, in order to facilitate the mere exposure effect (Emler et al., 2010). The objective of designing a service agent is to foster interaction with customers. Therefore, similarities such as human-like appearance or personality could be considered in robot design. Van et al. (2006) also supported with empirical evidence that the interface characters show similar body shape with users are perceived more credible and trustworthy than dissimilar one in the e-health context.

As a human, we enjoy the company of each other. We build teams (e.g. families, communities) and influence each other’s lives; we help each other because the fulfillment of the need to belong brings us happiness (Cacioppo & Patrick, 2008; Ryan & Deci, 2000). In the robotics field, research demonstrates people try to bond and affiliate with objects that are interactive and with basic social cues (e.g. speech) (Reeves & Nass, 1996; Nass & Moon, 2000). Therefore, it can be assumed that human apply similar criteria as they have in human-human relationship, to determine if they would like to interact with agent/robots. Based on the first two prerequisites (i.e. propinquity and similarity) from the need to belong theory, the following two hypotheses are therefore formulated. To facilitate the propinquity effect, a robot with physical embodiment (i.e. human-like robot and machine-like robot) is able to create more visual exposure effect than a virtual agent, even when the agent is not delivering verbal service. Additionally, similarity effect is applied when the robot looks similar to a human; therefore, the second hypothesis is developed.

H1: Consumers prefer a machine-like robot over a virtual agent. H2: Consumers prefer a human-like robot over a machine-like robot.

2.3.2. Social interaction

(24)

23

better relationship with others (Aronson et al., 2007). One of the main exchange principles outlined in SET literature is the ‘reciprocity rules’ (Cropanzano & Mitchell, 2005). Gouder (1960) distinguishes three different forms of reciprocity, and the first category ‘reciprocal interdependence’ is especially addressed in this paper. Reciprocal interdependence indicates the continuous interpersonal transactions, which means when a person actively delivers a favor; the counterpart would feel obliged to kindly repay the benefits (Gergen, 1969). Molm (1994) suggests reciprocal interdependence consequently decreases risk and enhanced cooperation among subjects. The exchange takes place when a person takes action, the other reciprocates. Once the exchange is initiated, reciprocity becomes a continuous process that each consequence of exchange self-reinforces the occurrence of the others. In the HRI literature, the rule of reciprocity is also applied, people feel the need to reciprocate when technology has done a favor for them (Fogg, 2002).

In addition, the ‘theory of mind’ indicates that humans are able to view other entities as intentional agents, whose behaviors are affected by states, beliefs, desires, and the knowledge that other humans wish, feel, know or believe something (Premach & Premach, 1995; Premach & Woodruff, 1978; Whiten, 1991). It has been seen as a fundamental prerequisite for human-human interactions. One of the terms representing the ‘theory of mind’, mind-reading, explains people are able to interpret the mental states of others. It can be assumed when the agents are capable of predicting what customers need, such as getting familiar with the hotel facilities and informing Wi-Fi password before they ask for it, customers will be more willing to interact with agents. Moreover, many researchers consider ‘theory of mind’ as a useful insight that can be implemented in agents (Peters, 2006), robots (Breazeal et al., 2004), or multi-agent systems (Marsella & Pynadath, 2005).

(25)

24

react favorably and exhibit reciprocity behavior to the agent if the service agent initiates the exchange cycle (i.e. active interaction). The following hypothesis is thereby developed.

H3: Consumers prefer active social interaction over passive social interaction.

2.3.3. Interaction effect between type of embodiment and social interaction

Besides two main effects of embodiment and social interaction, the possible interaction effect between these two characteristics needs to be investigated. There are both human and non-human characteristics combined in the agent/robot, which point out the conflicting cues might violate humans’ expectation. According to the ‘theory of category uncertainty’, the unease feeling towards robot arise when human feel ambiguous in distinguishing robot and human. The perceptual expectation is violated when robots look like a human, but do not act like a human. Besides, the concept of expectations is not new in marketing field. Consumer satisfaction literature considers product satisfaction as a function of consumer expectations and perceived performances (Solomon et al., 1985). The more negative inconsistency between expectations and performance, the more dissatisfaction consumer perceived (Churchill & Surprenant, 1982; Swan & Coombs, 1976). Similarly, the ‘role theory’ suggests satisfaction with service encounter depends on the congruence between perceived behavior and the behavior expected by consumers (Riordan et al, 1997). In addition, Evans (1963) indicates an ideal insurance agent meets the consumer expectations in terms of similarity, expertise, friendliness and personal interest. In the hotel setting, human service agents are generally expected to initiate active social interaction with customers, e.g. greeting people in a friendly way and engaging in asking feedback from customers. Based on the ‘theory of category uncertainty’ and ‘role theory’, it can be argued that a human-like robot, associating with a human service agent, is expected to perform human-like behavior, and the effect is stronger than a machine-like robot. The hypothesis is formulated as follows.

H4a: The preference of active social interaction is stronger for a human-like robot than for a machine-like robot.

(26)

25

2.3.4. Interaction effect between travel motives and type of embodiment

Researches regarding consumer’s traveling motivations will be elaborated, to understand the service preference and expectation of consumers. Previous research has concluded the value as a tradeoff between quality and price (Bolton & Drew, 1991). However, many scholars suggest that more aspects should be taken into account as consumer’s choices are influenced by multiple aspects of consumption value (Bolton & Drew, 1991; Grewal et al., 2003; Holbrook, 1994). There are two dimensions commonly used in consumer research literature, which are ‘utilitarian and hedonic value’ (Babin et a., 1994; Sheth, 1981). Utilitarian consumer behavior is illustrated as task-related and rational (Batra & Ahtola, 1991; Sherry, 1990). Consumers with utilitarian motives tend to demonstrate purchasing behavior in a deliberating and efficient way (Babin et al., 1994), also described as ‘problem solvers’ (Hirschman and Holbrook, 1982). On the other hand, consumers with hedonic motivations are illustrated as those who search for fun, fantasy, arousal, sensory, enjoyment, and emotional fulfillment from service experience (Hirschman and Holbrook, 1982).

(27)

26

motives; however, there is relatively lack of research apply the consumer motive in robotics field. Only one research conducted in Korea examines the distinct effects of hedonic and utilitarian robots, and the result suggests that people who have higher tendency to anthropomorphize also show more positive attitudes towards robot than those with low tendency to anthropomorphize (Lee et al., 2001).

It is reasonable to assume that a robot with anthropomorphic appearance (i.e. human-like robot in this study) is more attractive for hedonic travelers (i.e. holiday travelers), who expect to have an excited and enjoyable experience (e.g. family trip), than utilitarian travelers (i.e. business travelers), who value the service efficiency of service encounter and take the stay as the necessary part of subsistence during the trip. However, it is worth noticing that two motives can be presented on a continuum instead of two extreme states. For instance, utilitarian travelers (i.e. business travelers) might also show hedonic motive (e.g. excited about exploring a new place besides work) and vice versa. Two hypotheses concerning traveling motives are thereby formulated below.

H5a: Consumers with hedonic motive prefer a human-like robot over a machine-like robot. H5b: Consumers with utilitarian motive prefer a machine-like robot over a human-like robot.

Hypotheses for main effect

H1 Consumers prefer a machine-like robot over a virtual agent. H2 Consumersprefer a human-like robot over a machine-like robot.

H3 Consumersprefer active social interaction over passive social interaction.

Hypotheses for interaction effect

H4a The preference of active social interaction is stronger for a human-like robot than for a machine-like robot.

H4b The preference of passive social interaction is stronger for a machine-like robot than for a human-like robot.

H5a Consumers with hedonic motive prefer a human-like robot over a machine-like robot. H5b Consumers with utilitarian motive prefer a machine-like robot over a human-like

robot.

(28)

27

3. Research design

3.1.Methodology

In this paper, the attributes of a service agent/robot in hotel room will be analyzed to test consumer preference in adopting service agent/robot in hotel industry. In order to understand consumer preference on products or services of multi-attribute, Choice-Based Conjoint (CBC) analysis is suggested as the most appropriate method (Hagerty, 1895). Besides, CBC analysis is suitable for predictions about the diffusion of innovations or preference-based market segmentation (Eggers, 2011). Respondents will be invited to repeatedly choose their most preferred hotel room between 3 sets of alternative, and then the optimal combination will be determined. Besides, two surveys with scenario of holiday and business trips respectively will be designed for testing moderator of hedonic and utilitarian traveling motives. Thirdly, the interaction effect between type of service agent and social interaction will be examined with the collected data. Lastly, several segments will be created to cluster respondents based on their preferences.

3.2.Study design

In this section, the selected attributes and levels of service agent in the hotel room will be determined, which are shown in table 5. In order to minimize the standard error, the number of levels for each attributes should be kept fewer and precisely. Furthermore, the number of levels should be as balance as possible to avoid number-of-levels effect. The effect occurs when the number of levels is not distributed evenly across attributes, thus resulting in higher importance of attributes with more levels. Lastly, each attribute should be mutually exclusive of each other, so every combination with different levels per attribute is possible.

Attributes Level 1 Level 2 Level 3

Type of

embodime nt

Virtual agent Machine-like robot Human-like robot

Social interactio n

Passive Active

price €100 €110 €120

(29)

28

3.3.Experimental design

To conduct an effective experimental design, the attribute ‘type of embodiment’ will be visualized through the pictures, while the social interaction and price will be explained individually in the choice sets. The full factorial design with all possible attribute level combinations will be 3x2x3=18, which leads to 18!

(18−3)!∙3!. = 816 possible choice sets with 3

stimuli per choice set. To reduce the number of possible combinations, a random sequence choice design that shows 12 choice sets with three hotel room alternatives each to respondents is applied. In addition, the no-choice option (i.e. ‘would you actually buy your preferred choice with indicated price if it is available?’) is included as a separate question after each choice task to increase realism without losing information about respondents’ relative preferences on other alternatives.

Two studies were created following the same choice design while controlling moderating preference with hedonic and utilitarian traveling motives. For the hedonic motive, respondents were asked to imagine a situation in which they are going on a holiday trip and staying at a luxurious hotel which provides a service agent that deliver verbal customer service. For the utilitarian motive, respondents were asked to imagine a situation in which they are going on a business trip and staying at a business hotel which provides the same service as the hedonic scenario. The example conversations are different in order to facilitate imagination in holiday and business trips. Introduction pages and examples of conjoint choice sets from two surveys are displayed in appendix A.

(30)

29

For the second variable ‘frequency of staying at hotel’, the scale was developed by Akbaba (2006). Respondents were asked to indicate their frequency of staying at hotels from ‘less than once a year’, ‘once a year’, ‘twice a year’ to ‘five times or more a year’.

Lastly, some socio-demographic (i.e. age, gender, education, income, occupation, nationality) information was asked. The respondent behavioral engagement during the survey is tested by one additional question inserted in the section of attitude towards robot. Respondents had to answer the specific number (i.e. ‘please indicate 2 for this question’); therefore, people had to read the question carefully. Respondents who did not pass the attention check are excluded from the data set.

3.4.Data collection

To ensure enough data are collected, a rule of thumb was used to determine the appropriate sample size. There are (3+2+3)-3=5 parameters in the model, which require 5 observations to estimate each parameter for each scenario. Therefore, 5*5=25 observations are required for one scenario. In total, 50 respondents are sufficient.

The website prederencelab.com, which is tailor-made for CBC analysis, is used to design the surveys. Two surveys were combined into one entry link, and distributed through Facebook, Line and WhatsApp application. Therefore, respondents are randomly allocated into one scenario. The between-subjects design has two main benefits that strengthen internal validity. Firstly, every respondent is only accessible to one scenario, which people are less likely to suppose researcher’s intention that confounds the results. Furthermore, the learning effect of filling out both hedonic and utilitarian surveys is eliminated; otherwise, people familiar with one testing environment than another will also confounds the result. People with diverse background were approached to vary the participant portfolio. In the end, 107 respondents completed the hedonic survey and 112 completed the utilitarian survey.

3.5.Modeling approach

3.5.1. Main effect – aggregate model – model 1

(31)

30

embodiment’ and ‘social interaction’ are nominal. The model includes the largest amount of parameters that estimate each attribute levels individually. In addition, the linear or ideal point model can also be applied for the ‘price’ attribute. These two preference functions are more simple than part-worth since it considers less parameters. In order to determine how the ‘price’ attribute should be treated and which type of models to be used, the model fit will be compared. Ultimately, the model fitting the best would be chosen as the ‘aggregate model’, which will be used to analyze the main effects and compare them with the hypotheses.

3.5.2. Interaction effect – ‘type of embodiment’ and ‘social interaction’ – model 2

Secondly, the interaction effect between ‘type of embodiment’ and ‘social interaction’ will be investigated. To test such an effect, new variables need to be created by multiplying effect-coded attribute levels of ‘embodiment’ with other levels of ‘social interaction’. In this case, the interaction between two attributes is recoded into two new variables (i.e. ‘machine-like robot x passive interaction’, ‘virtual agent x passive interaction.’), whereas the effect coding of human-like robot and passive interaction is used as baseline. Finally, these new variables are included in the model 2, to check if the interaction effect occurs.

3.5.3. Interaction effect – ‘travel motives’ and ‘type of embodiment’– model 3

Lastly, model 3 is created to test whether the interaction between ‘travel motives’ and ‘type of embodiment’ occurs. Two new variables (i.e. human-like robot x motive; virtual agent x motive) indicating the interaction effects between ‘type of embodiment’ and ‘traveling motives’ are firstly created by multiplying the effect-coded attribute levels of ‘embodiment’ with ‘motives’. Then, model 3 is developed by including two interaction variables and compared with aggregate model to find out whether the model fit improved.

3.6.Assessment of model fit

Model fit is used to compare whether one model fit significantly better than the other. There are several ways to evaluate the model fit. First of all, likelihood ratio (LR) test is used to test whether the estimated model parameters are significantly different from zero. The test is frequently used to test the significance when including many explanatory variables (Crichton, 2001). In order to conduct the LR test, chi-squared statistic is calculated by the function 𝑥2 ₌

(32)

31

distribution equals the difference in number of parameters (npar). The p-value is examined to determine significance. Furthermore, AIC, AIC3, BIC, CAIC are four important information criteria that are calculated based on log-likelihood test while penalizing for number of parameters. It is suggested to choose model with minimum value within the same criteria (Vermunt and Magidson, 2015). There is no best option for those four criteria, while they are suitable for different situation. AIC and AIC3 tend to favor more complex models with small penalty. On the other hand, BIC and CAIC penalize for higher complexity (i.e. more latent classes) and are especially preferred for large sample sizes. Therefore, BIC and CAIC would be suitable criteria to assess model fit in this research due to the sample size with lots of observed choices.

In addition, the R² and McFadden’s adjusted R² are an alternative to evaluate model fit. They are calculated with the following function, 𝑅2 _{= 1 −}𝐿𝐿(𝛽)

𝐿𝐿(0) and 𝑅

2_{𝑎𝑑𝑗 = 1 −}𝐿𝐿(𝛽)−𝑛𝑝𝑎𝑟 𝐿𝐿(0) ,

where LL(β), LL(0) and npar represents the same as the likelihood ratio test. According to MacFadden (1979), adjusted R² within the range of 0.2 and 0.4 can be considered acceptable. Thirdly, the hit rate indicating the percentage of the holdout choice set that are predicted correctly is another way to test how good the model can predict respondents’ actual choices.

(33)

32

4. Results

The result of the CBC analysis will be revealed in this section, to provide an overview on consumer preferences for service agent in hotel industry. To begin with, descriptive statistics will be presented to take a glimpse at respondents’ characteristics. Furthermore, three models will be created via CBC analysis. Model 1 is an aggregate model that is created to investigate the main effect. Model 2 is developed to test if there is interaction between ‘type of embodiment’ and ‘social interaction’. Model 3 accommodates the interaction effect between ‘motives’ and ‘type of embodiment’. Lastly, the latent-class analysis will be conducted to cluster respondents that are homogenous regarding their preference. The segmentation will be used for managerial implication.

4.1.Data preparation

Variables Categories Frequency percentage

Age 0-20 20-30 35-60 >60 years old 17 133 13 1 10.4 81.1 7.9 0.6 Gender Male Female 65 99 39.6 60.4 Income €0-€25,000 €25,000-€50,000 €50,000-€75,000 €75,000-€100,000 more than €100,000 126 22 5 1 10 76.8 13.4 3.0 0.6 6.1 Occupation Student Non-student 114 50 69.5 30.5 Nationality Asian European Others 116 38 10 70 23.2 6.1

Level of education Non-university degree University degree

16 148

9.8 90.2

Frequency of staying at hotel Less than once a year Once a year

Twice a year Three times a year Four times a year Five times or more a year

19 34 30 32 10 39 11.6 20.7 18.3 19.5 6.1 23.8

Traveling motives Hedonic Utilitarian

79 85

48.2 51.8

Table 6: demographics of 164 respondents

(34)

33

variables are organized into lesser categories to prepare for further segmentation. For ‘nationality’, respondents are grouped into ‘Asian’, ‘Europe’, and ‘Other’. For “age”, they are regrouped into ‘0-20’, ’20-35’, ’35-60, and ‘above 60’ years old. For “level of education”, they are divided into ‘non-university degree’ and ‘university degree’. Income, frequency of staying at hotels and attitude towards robot remain in the same order. After the data cleaning, the demographic statistics combining two datasets are summed up in table 6. For the ‘nationality’, there are 70% Asian, 23.2% European and 6.1% participants from the other continents. After that, data from conjoint choice sets need to be combined with the other control variables in order to perform analysis in Latent Gold program.

Finally, the variable ‘negative attitude towards robot’ needs to be addressed by conducting factor and reliability analysis. There are 14 items in the survey concerning attitude towards robot. The factor analysis is firstly performed, which loaded to three factors. The result shows KMO measure of sampling adequacy is 0.749, implying factor analysis is appropriate since it should be above 0.5 (Kaiser, 1974). Sample adequacy predicts if data are likely to factor well, based on correlation and partial correlation. In addition, the Bartlett’s Test of Sphericity shows the significance of 0.000, indicating the factor is significant. Therefore, the null hypothesis, ‘the variables are uncorrelated’ can be rejected. Thirdly, the result of Principal Component Analysis (PCA) shows the eigenvalue of 4.718, 1.928 and 1.283 for the first three components, which are sufficient above 1. The accumulative explained variance for three components is 56.639, which is above 50% and above 5% each. The rotated component matrix also shows that the loadings for three loadings are above 0.5, thus no items need to be deleted. Next, the reliability test is performed to test if the theoretical factor consisting three marker items is also strong enough in this research. For the variable ‘negative attitude towards robot’, the reliability statistics shows the Cronbach’s alpha of 0.826, implying the items are highly correlated. The value is sufficient above 0.6 (Peterson, 1994), and the Cronbach’s alpha did not increase if item deleted. Finally, three variables indicating the negative attitude towards ‘situation of interaction with robots’ (factor scores=3.47), ‘social influence of robots (2.87)’ and ‘emotions in interaction with robots (4.08)’ are created. The higher the value in the Likert scale, the more negative the attitude towards the robot.

4.2.CBC analysis – main effects

(35)

34

effect of traveling motivations will be tested. The interaction effect of two attributes will be investigated and compared to the hypotheses.

4.2.1. Model selection

To begin with, the part-worth model can be created by defining three attributes and non-option as nominal, therefore part-worth utilities can be estimated for all attributes. However, it can be argued that the price attribute can be specified as linear, thus the second model is developed. In order to determine what type of model should be used, the Likelihood ratio test is conducted to see whether adding more parameter significantly improves fit. The result (Chiaq-statistics=1.2196, df=1, p>.05) shows that the part-worth model with one more parameter does not significantly increase model fit, thus the price function should be defined as linear. The model 2 solution is further confirmed by the lower BIC (5773), CAIC (5748) as shown in table 7. Next, the predictive validity of model is compared with the NULL model before further interpretation.

Model 1: part-worth model Model 2: price as linear model

Npar 6 5 LL -2873.46 -2874.07 BIC 5777.5207 5773.6404 CAIC 5783.5207 5748.1411 R² 0.1851 0.1849 Adjusted R² 0.1834 0.1835 Hit rate 0.670732 0.670732

P-value 6.1e-1094** 2.0e-1093**

** p < .01

Table 7: comparison on goodness of fit - price as part-worth or linear model

4.2.2. Validity

(36)

35

In sum, model 2 is valid according to likelihood ratio test and hit rate, thus the estimation result can be used for interpretation.

4.2.3. Interpretation of the main effects

Attributes Levels Utility Standard error

Wald p-value Mean Type of embodiment human-like robot -0.0453 0.0317 34.6375 3.00E-08** -0.0453

virtual agent -0.1298 0.0323 -0.1298

machine-like robot

0.1752 0.0306 0.1752

Social interaction Passive 0.2754 0.0238 134.1785 5.00E-31** 0.2754

active -0.2754 0.0238 -0.2754

Price -0.5854 0.0282 430.042 1.60E-95** -0.5854

Non option -2.2807 0.0765 889.3317 2.00E-195** -2.2807 **P<.01

Table 8: estimation result of the main effect from aggregate model

The result for the aggregate model is shown in the table 8. All attributes turn out to be highly significant. Furthermore, the t test is calculated to investigate if all attribute levels significantly different from one another. In terms of ‘type of embodiment’, the estimates imply the part-worth utilities such that respondents prefer the machine-like robot (U = 0.1752) much more than human-like robot (U = -0.0453) and virtual agent (U = -0.1298). Besides, t test indicates the difference between machine-like robot and virtual (t=6.8550, p=0.0001), human-like and machine-like robot (t=5.0046, p=0.0001) are significant. The finding supports H1: Consumers prefer a machine-like robot over a virtual agent. However, H2:

Consumers prefer a human-like robot over a machine-like robot is not supported. Secondly,

for the attribute ‘social interaction’, respondents generally prefer passive (U = 0.2754) over active (U = -0.2754). T test indicates there are significant difference between these two levels, i.e. passive and active interaction (t=16.3645, p=0.0001). which is contradicted towards H3:

Consumer prefer active social interaction over passive social interaction. Thirdly, the price

(37)

36

4.3.CBC analysis – interaction effect of type of embodiment and social interaction

4.3.1. Model selection

Model 2: aggregate model Model 3: including interaction effect Improve fit?

Npar 5 7 LL -2874.07 -2873.33 No BIC 5773.6404 5782.3538 No CAIC 5748.1411 5789.3538 No R² 0.1849 0.1851 Yes Adjusted R² 0.1835 0.1832 No Hit rate 0.670732 0.670732 No

P-value 2.0e-1093** 1.1e-1094**

** p < .01

Table 9: comparison on goodness of fit - aggregate or interaction-included model

In order to examine the interaction effect between ‘type of embodiment’ and ‘social interaction’, two new interaction variables (i.e. human-like robot x passive interaction, virtual agent x passive interaction) are included in model 3. Firstly, the likelihood ratio test (Chisq=1.4864, p>.05) indicates model 3 including interaction effect is not significantly different than aggregate model in terms of model fit. Secondly, the finding is further confirmed by the information criteria (i.e. BIC: 5782 > 5773 and CAIC: 5789 > 5748) of the model 3 scores worse than the aggregate model without interaction effect. In sum, model 3 including the interaction between ‘type of embodiment’ and ‘social interaction’ does not make better prediction than aggregate model without interaction. Therefore, the H4a: The

(38)

37

4.4.CBC analysis – interaction effect of traveling motives and type of embodiment

In this section, the interaction effect of traveling motives and type of embodiment will be analyzed. The hypotheses will be further compared with the estimation result.

4.4.1. Interpretation of the moderating effect

Attributes Utility Wald p-value Mean Type of embodiment human-like robot 0,0018 5,4056 0,067 0,0018 virtual agent -0,092 -0,092 machine-like robot 0,0903 0,0903 Social interaction passive 0,2766 135,2234 3,00E-31** 0,2766 active -0,2766 -0,2766 Room price -0,5899 433,511 2,80E-96** -0,5899 None_option -2,2416 763,3687 5,00E-168** -2,2416 Interaction variables

Motive x machine-like robot 0,2742 2,7161 9,90E-02** 0,2742

Motive x virtual agent 0,0193 0,0355 0,85 0,0193

Motive x human-like robot 0.7065 Reference level

** p < .01, * p < 0.05

Table 10: estimation of model with interaction effect of travel motives

In order to test the interaction effect of traveling motives, two new variables were created by multiplying ‘motive’ and the effect-coding of machine-like robot and virtual agent. The interaction effect between travel motive and effect-coding of human-like robot is used as baseline. As can be seen from the table 10, the interaction effect between traveling motive and machine-like robot is significant (p<.01, U=0.2742), which indicates the positive effect of a machine-like robot is stronger when participants take a business trip (utilitarian scenario) instead of a holiday trip (hedonic scenario) compared to a human-like robot, and vice versa. Therefore, H5b: Consumers with utilitarian motive prefer a machine-like robot over a

(39)

38

4.5.Hypotheses overview

So far, the main and interaction effects have been investigated, an overview for hypotheses is shown in the following (table 11).

Hypotheses for main effect Result

H1 Consumers prefer a machine-like robot over a virtual agent. Supported H2 Consumersprefer a human-like robot over a machine-like robot. Rejected H3 Consumersprefer active social interaction over passive social interaction. Rejected

Hypotheses for interaction effect

H4a The preference of active social interaction is stronger for a human-like robot than for a machine-like robot.

Rejected

H4b The preference of passive social interaction is stronger for a machine-like robot than for a human-like robot.

Rejected

H5a Consumers with hedonic motive prefer a human-like robot over a machine-like robot.

Supported

H5b Consumers with utilitarian motive prefer a machine-like robot over a human-like robot.

Supported

Table 11: Hypotheses overview

4.6.Latent class analysis – segmentation

With the aggregate model developed before, we assume that there is no preference heterogeneity across consumers. In this section, the segments will be created based on the three attributes and control variables (i.e. age, gender, income, education, occupation, nationality, frequency of staying in hotel and attitude towards robot). Firstly, the model includes all control variables as covariate to improve classification. Next, the optimal number of segments will be determined by estimating models from 1 to 10 segments multiple times. The model criteria for these 10 models are summarized in table 12.

LL BIC(LL) CAIC(LL) Npar L² df p-value Class.Err. Model1 1-Class Choice -2874,07 5773,64 5778,64 5 5748,141 159 2,0e-1093 0,00E+00 Model2 2-Class Choice -2616,59 5289,283 5300,283 11 5233,184 153 2,2e-989 2,30E-02 Model3 3-Class Choice -2463,46 5013,612 5030,612 17 4926,915 147 2,1e-929 3,61E-02 Model4 4-Class Choice -2354,04 4825,386 4848,386 23 4708,089 141 7,1e-888 3,18E-02 Model5 5-Class Choice -2273,06 4694,013 4723,013 29 4546,117 135 2,5e-858 4,14E-02 Model6 6-Class Choice -2198,12 4574,734 4609,734 35 4396,239 129 2,5e-831 3,46E-02 Model7 7-Class Choice -2148,13 4505,354 4546,354 41 4296,259 123 7,4e-815 0,0469 Model8 8-Class Choice -2106,82 4453,339 4500,339 47 4213,645 117 4,5e-802 0,0326 Model9 9-Class Choice -2071,77 4413,829 4466,829 53 4143,536 111 5,8e-792 0,0276 Model10 10-Class Choice -2036,05 4372,988 4431,988 59 4072,096 105 1,3e-781 0,0319

(40)

39

Figure 13: Select number of segments based on BIC and CAIC

The model with the best fit will be selected according to information criteria, classification error and interpretability. As can be observed from figure 13, both lines become flat from the solution of 5 segments, which means the difference of BIC and CAIC between segments becomes smaller after solution of 5 segments. Therefore, the solution of 4 segments, one point before the elbow point, is chosen.

Attributes Segment 1 Segment 2 Segment 3 Segment 4 p-value p-value Type of embodiment

human-like robot -0,6588 -0,6189 0,9058 0,1045 3,50E-50** 8,60E-49** virtual agent 0,3008 0,2074 -1,1836 -0,3562

machine-like robot 0,358 0,4115 0,2779 0,2517

Social interaction

passive -0,1209 1,592 -0,0776 0,4209 5,80E-54** 3,20E-53** active 0,1209 -1,592 0,0776 -0,4209 Room price -0,8991 -1,0216 -0,3714 -0,9105 8,30E-92** 2,40E-07** None_option -4,3341 -2,2179 -1,8754 -0,3382 6,10E-79** 2,80E-33** Covariate Att_SI -0,2778 0,4944 -0,3331 0,1166 0,015* ** p < .01, * p < 0.05

Table 14: the characteristics for 4 segments

The estimation for 4 segments is shown in table 14. The second column of p-value shows a significant result for three attributes (p<.01), which means consumer preferences for these 4 segments do significantly differ from one another. Furthermore, only the covariate ‘negative

Measuring consumer preferences for automated agents in a hotel context

Measuring consumer preferences for

automated agents in a hotel context

Differences between hedonic and utilitarian travel motives

By

Shu Han, Chuang

Measuring consumer preferences for

automated agents in a hotel context

Differences between hedonic and utilitarian travel motives

Master Thesis

MSc Marketing Intelligence

University of Groningen, Faculty of Economics and Business

1

Supervisor: Dr. Jenny van Doorn

2

supervisor: Dr. Lara Lobschat

completion date: 26

June, 2017

Author: Shu Han, Chuang

Student number: 3077918

Address: Duindoornstraat 45, 9741NN, Groningen

Phone number: +31 63392 4760

Abstract

Preface and acknowledgement

Table of Contents

1. Introduction

2. Theoretical framework

3. Research design

4. Results