Can we get used to robots? Exploring the uncanny valley effect over time

(1)

Can we get used to robots?

Exploring the uncanny valley effect over time

Project: Thesis MSc Marketing Intelligence

Faculty: Economics and Business

First supervisor: dr. J. van Doorn

Second supervisor: prof. dr. ir. K. van Ittersum

Date: 15-01-2018

Student: Lotte Sarah Logher

(2)

2

Abstract

(3)

3

1. Introduction

In recent years, robots have become a larger part of everyday life. Nowadays, they are manufactured to help us solve problems and perform better, to keep us safer and even to perform tasks that used to be performed by human beings. Robots have over the years proven to be capable, productive and able to perform valuable work (Lawton, 2017). Research suggests that levels of automated social presence will only increase and are affecting customer service perceptions and outcomes (Van Doorn et al., 2017). Social presence in this case is defined as the sense of being with another, either a human presence or an automated presence (Biocca, Harms and Burgoon 2003; Heeter 1992). These changes lead to an ever increasing role of robotics which could potentially revolutionize the way we live (Del Prado, 2015).

In several industries, innovations in robotics are already implemented. These implementations are the most apparent when it concerns physically embodied robots. As opposed to virtual robots, embodied robots are presented in a tangible and visible form. These embodied service robots are used to perform tasks more effectively and efficiently than human beings. Especially in service settings such as healthcare and customer service, human beings are more and more confronted with an increase in robotics. Examples of embodied robots in healthcare settings are the Xenex robot which is able to quickly and effectively disinfect any space within a healthcare facility, a surgical robot which enhances a surgeons vision, precision and control during an operation and the TUG robot that is able to carry around sensitive materials like medication and laboratory specimens around the clock in a hospital (Meskó, 2017).

Examples of robots in customer service are the social humanoid robot Pepper that can interact with customers in 20 different languages (Meskó, 2017), the robot Tally that can roll around retail stores autonomously to find products that are misplaced, mispriced or low in stock (Baird, 2017) and robots performing basic tasks in cafés or restaurants so human employees are allowed more time to engage in social encounters with customers to provide a better customer experience (Hu, 2017). Experts expect that robots will continue to take over more basic and easily automated tasks across industries. This could pave the way for a more efficient and productive workforce (Hu, 2017). For example, robots make it possible to provide services immediately and 24 hours per day, which is impossible for human employees (Hyken, 2017).

With robots being infused into service contexts more and more, the question rises to what extent people are comfortable with this increasing robotic presence in their everyday lives and how they will react to this change. Attitudes towards machine-like robots in assisting roles seem to be mostly positive, but people seem to feel negatively towards robots once they are more human-like and able to do tasks by themselves instead of merely assisting a human being. This second kind of robot tends to induce a feeling of discomfort (Burns, 2016). These outcomes are in line with the 2015 report on autonomous systems by the European Commission, which presented Europeans opinions towards robots and the future use of robots, autonomous cars and drones. This report states that Europeans overall view robots positively and are confident that they could perform jobs that might be too hard or dangerous for humans. However, the study also found that almost 90 percent of the respondents has the opinion that advanced robots should be carefully managed. The report concludes that mainly older people are comfortable with mechanical robots, but perceive less affinity towards more humanoid robots (Burns, 2016).

(5)

5 more humanlike robots. This level of affinity increases until the uncanny valley is reached; very high levels of human likeness can cause this feeling of affinity to change into an eerie sensation. This might happen when someone realizes that the features that at first sight very closely resembled a human being, are in fact artificial. The result is a non-linear effect of the human-likeness of a robot on the affinity that it induces in people. Mori (2012) therefore proposed that choosing a moderate degree of human likeness would be most preferable in creating robots which would cause a considerable sense of affinity without scaring people.

Right now, people are still figuring out how to treat robots, but it is expected that over time norms about interaction with robots will be developed within human culture. By that time, robots might have become a part of daily life and evoke more pleasant associations (Burns, 2016). For example, robots performing domestic tasks at home might be considered as familiar as a dishwasher (Lin, 2016). So, even though people might have a negative first reaction to a robot, this effect might decrease or even disappear over time as a person gets more used to the robot after several encounters. At this point in time, only one experiment has been conducted to investigate the uncanny valley effect over time. Zlotwosky e.a. (2015) have studied how people react to two types of embodied robots over multiple encounters. They found that the merely being confronted with these robots repeatedly can reduce the eerie feeling that the robots initially evoked.

However, even though some research has been performed on the uncanny valley effect over time, current literature is lacking in this area. First of all, little research has been performed on how different kinds of robots are perceived in comparison to humans in a service context. By comparing the effects of robots to the current benchmark, which are humans as service employees, more insight can be provided in how people perceive robots in comparison to human service employees. Second, Zlotwosky e.a. (2015) found that the reduction in eeriness was the same for two different robots even though the robots initially evoked different levels of eeriness. However, the question remains how these changes compare to encountering a human several times. Further research into this topic could show whether or not the gap in perception of robots versus the perception of humans can become smaller, thus implying that over time people could get used to robots in service settings. Third, no research has been done yet on possible lagged effects that could occur over time when it comes to repeated encounters with robots. Previous research suggests that the results of previous stimuli encounters can be subject to delay, which could mean that the results of following encounters are affected by the previous encounter (Bornstein, 1989; Stang, 1975).

(6)

6 This paper will focus on two specific service outcomes. These service outcomes are compliance and loyalty. Both compliance (Reingen, 1978; Hausman, 2004) and loyalty (Yim e.a., 2008) have over time proven to be important aspects in different service contexts. First, a large component of the service industry is the mere interaction of service employees with customers to actually provide them with good customer service. One trait that is of importance for service employees in a variety of contexts is their ability to obtain compliance. Compliance in these cases is defined as the extent to which customers follow orders of a service provider (Hausman, 2004). It plays a role in, for example, the context of buying a new product or completing a survey (Reingen, 1978). But also healthcare is a service context where compliance can play a large role, for example in encouraging compliance with medical advice (Hausman, 2004). The second service outcome that is known to be an important indicator of a customer’s perception of a service, is loyalty. According to Yim e.a. (2008), building customer loyalty is a key priority for service managers. The reason for this priority is that stable customer relationships lead to higher firm profits. One key determinant of customer loyalty is repurchase intention (Dong, 2011). A second important driver of customer loyalty is recommendation intention (Ying and Meng, 2009).

The research questions that will be answered in this paper are the following:

 How does the perceived eeriness of different kinds of robot service providers change after several encounters in comparison to human service providers?

 What is the effect of different robot service providers on the service outcomes compliance and loyalty in comparison to human service providers?

This paper is divided in four parts. First of all, the existing knowledge concerning the research question at hand will be discussed to provide a theoretical framework. Secondly, the method of data collection and analysis will presented. Thirdly, the results of the study will be described and finally the conclusions and managerial implications will be discussed.

2. Theoretical background

2.1 The uncanny valley effect

(7)

7 Figure 1: The uncanny valley effect (Mori, 2012)

Recent studies have tried to pinpoint what exactly causes this uncanny feeling as a robot comes closer to resembling and actual human being. Mori himself (1970) initially proposed that movement could have a large part in causing the feeling of eeriness, as this could lead to a contrast between a visually realistic robot and tactile imperfections. However, over time other scholars have aimed to add insights to Mori’s initial theory. A short summary of the following findings can be found in Table 1.

Study Findings

Mori (1970) The uncanny feeling is a result of a contrast between visual realism and tactile imperfections (Violation of Expectation).

Jentsch (1906), Burleigh & Schoenherr (2015)

The uncanny feeling is a result of people’s categorical perception of a robots human-likeness, instead of affective rating causing an eerie sensation (Categorical Uncertainty).

Hanson (2005), MacDorman & Ishiguro (2006)

The uncanny feeling is a result of faces not meeting aesthetic standards, while people are highly sensitive to facial aesthetics (Evolutionary Aesthetics).

Ho e.a. (2008),

MacDorman & Ishiguro (2006)

The uncanny feeling is a result of faces reminding people of their own mortality, which provokes fear of death (Mortality Salience).

MacDorman e.a. (2009), MacDorman & Ishiguro (2006)

The uncanny feeling is a result of faces that indicate a heightened risk for transmissible diseases, which in turn triggers a mechanism for pathogen avoidance which leads to a disgusted sensation (Pathogen Avoidance). Gray & Wegner (2012) The uncanny feeling is a result of robots reminding people of their

identity as robots as opposed to humans, which leads to the question what the definition of being human is (Mind Perception).

Stein & Ohler (2016) The uncanny feeling is a result of the attribution of emotions and social cognition to a robot.

Table 1: Overview of uncanny valley findings

(8)

8 propose that each person attempts to segment these aspects into “human” and “non-human” categories. When one or more aspects of the robot do not specifically belong to one of the categories, this leads to a strong negative response. They found that a person’s category structures and the frequency of exemplars of these categories retained in memory affect the level of eeriness someone feels towards a robot. This theory is closely related to Jentsch’ cognitive dissonance theory, which claims that once a stimulus is not in line with someone’s expectations based on perceptual cues and former experiences, this leads to an emotional uneasiness or eerie sensation (Jentsch, 1906).

Second, there is a theory called Evolutionary Aesthetics (Hanson, 2005; MacDorman & Ishiguro, 2006), which states that humans are very sensitive to facial aesthetics. Hanson (2005) explains that this is the result of selection pressures for certain physical attributes that could health. When confronted with a stimulus with an uncanny face that fails to meet these aesthetics standards, a sensation of eeriness is evoked. This sensation is more the result of low attractiveness instead of the result of not being realistic enough (Hanson, 2005).

Third, Ho e.a. (2008) as well as MacDorman & Ishiguro (2006) proposed a theory called the Mortality Salience. This theory states that the uncanny feeling found in the uncanny valley is the result of humanlike faces reminding people of their own mortality. This in turn can lead to a fear of death, evoking the eerie sensation, because it triggers a defence mechanism in people (Ho e.a., 2008). Also, these kinds of robots can elicit the feeling of a threat, as people might feel that this humanlike replica might replace a human, which induces the feeling of bodily control (Ho e.a., 2008).

Fourth, the uncanny feeling could be a result of a humanlike face indicating a high risk of disease. This theory is called the Pathogen Avoidance theory (MacDorman e.a., 2009; MacDorman & Ishiguro, 2006). It argues that a humanlike face might elicit a feeling of disgust, because the imperfections are interpreted as indications of disease. This in turn triggers the pathogen avoidance mechanism, leading to the disgusted feeling. This theory thus bases the uncanny feeling on disgust rather than on fear. Fifth, Gray & Wegner (2012) hypothesized a theory called Mind Perception. This states that the uncanny feeling is the result of a robot reminding people of their own humanity. It leads to people comparing their own humanity as opposed to the robots identity. The fact that a robot can so closely resemble humans leads to the question on what bases people actually perceive other human beings. It leads to questioning what attributes actually define being a human. As these humanlike robots resemble humans closely, they trigger the attributions of human minds. This attribution of human experience to a robot is the cause of the uncanny feeling (Gray & Wegner, 2012).

Finally, Stein and Ohler (2016) found more recently that the eerie feeling can be evoked by the attribution of emotions and social cognition, rather than a robot being a closer physical resemblance of a human. If a robot was perceived as autonomous artificial intelligence, people perceived significantly higher levels of eeriness when compared to scripted robots.

(9)

9 2.2 Habituation in psychology

Until recently, little scholars have shown interest in the results of the uncanny valley effect over time. In psychology literature however, several theories have been developed when it comes to people being confronted with stimuli repeatedly. The reasoning behind this interest in multiple confrontations is that the ability of human beings to quickly classify a stimulus as “positive” or “negative” is a very vital psychological ability, which is inherently linked to our chances of survival. This system of classification therefore works very fast and automatically. However, repeated exposure to a stimulus could change this process and therefore a lot of research has been done in the area of classifying stimuli and how people react to stimuli over time (Dijksterhuis and Smith, 2002).

Repeated encounters with a stimulus often lead to habituation, as habits are developed when someone repeatedly responds to the same stimulus in a similar context. This leads to the forming of associations between the stimulus and a response. Once a habit is formed, a context can directly activate the associated response from someone’s memory. The initial response to a stimulus is quite important for habituation to occur. Initially, people’s behaviour often reflects a desired goal or outcome. When the desired outcome is met, an intention is formed to repeat this response in a future situation. While forming a habit like this, a person’s behaviour gradually shifts from goal-oriented to habitual (Wood and Neal, 2009). However, not all habits are formed in this conscious manner. According to Wood and Neal (2007), habits can also be formed based on automatic thought processes that consist of multiple separable features that can occur in different combinations. These kinds of habits are categorized as “goal-independent” as they do not depend on reaching a goal to occur (Wood and Neal, 2007).

There are two theories related to repeated encounters and habituation that are expected to influence the effect that repeated encounters will have on the perception of robots. The first theory claims that just exposing an individual to a stimulus several times instead of only once is considered to be a key determinant of affective response (Burleigh and Schoenherr, 2015). This is explained by the exposure effect, where mere exposure can lead to an increase in affective response (Bornstein, 1989). This theory was first introduced in 1968 by Zajonc, who stated that repeated exposure to a stimulus will lead to an increase in positive affect towards this stimulus. This relationship between exposure and affect has over time been confirmed to be a robust and reliable phenomenon for several stimuli in different settings (Bornstein, 1989). From this can be concluded that, when taking robots into account, the mere exposure effect implies that repeated encounters with a robot lead to an increase in positive affect towards the robot, even though the robot might initially evoke negative emotions. According to the theory, this negative initial perception could be reduced by merely increasing people’s familiarity with robots.

(10)

10 habituation process comes into play in these cases of extreme emotional stimuli to guide behaviour (Neal e.a., 2009). When linking this theory to encountering robots that evoke a strong and negative initial response, this means that simply being confronted with robots several times can reduce the initial reaction to a less strong negative feeling.

Both these theories are expected to come into play when investigating the uncanny valley effect over repeated encounters. For a robot that closely resembles a human, the uncanny valley effect would predict a strong negative perception. However, according to the affective habituation process, this outcome could be weakened by exposing the customer to the robot more than once and, according to the mere exposure effect, an initial neutral or even negative perception could turn into a positive perception after several exposures.

2.3 Uncanny valley in repeated encounters

A recent study by Zlotowsky et al. (2015) has found that perceived robot eeriness can be reduced by three encounters with the same robot. The reasoning behind this experiment was the researchers’ assumption that the high level of eeriness induced by a humanoid robot might be the result of the novelty of the stimulus. These feelings could diminish after it has become clear that the stimulus is not harmful (Lee, 2001). This would imply that the uncanny valley effect could decrease after several encounters with the stimulus. The researchers investigated two types of robots; an android that very closely resembled a human being and a machinelike robot with just a few humanlike features like arms and a head. These two robots were presented to participants during repeated interactions to investigate the possible changes in perception after several encounters. In line with the uncanny valley, they found that the android was rated as more eerie than the machinelike robot. Furthermore, they found that repeated interactions significantly reduced the level of perceived eeriness that the robots induced, irrespective of their embodiment. Also, this reduction was roughly similar for both robots, even though the initial levels of eeriness were different. The gap between the machinelike and humanlike robot stayed relatively constant over the three encounters. Based on these findings the researchers hypothesized a new graph that resembles a similar reduction in eeriness for robots irrespective of their embodiment.

Three explanations were provided to explain the findings. First of all, the affective habituation effect (Dijksterhuis and Smith, 2002) was suspected to influence the perception of the android, which was initially perceived as extremely eerie. As the android’s perceived eeriness decreased after several encounters, this process could have weakened the initial reaction towards the robot. Second, the mere exposure effect (Burleigh and Schoenherr, 2015) was suspected to influence the perception of the machinelike robot, which was initially perceived somewhat neutral. As this reaction became more favourable after several encounters, the additional encounters were suspected to have increased affection towards the robot. A third explanation lies in the novelty of the robots as stimuli, which could result in extreme first reactions that wear off over time (Lee, 2001). It could become easier to process these stimuli after more encounters as its appearance becomes more familiar (Bornstein and D’Agostino, 1994).

Independently of the explanation of the phenomenon, it can be concluded that people are able to get used to the unfamiliar appearances of different kinds of robots, as Zlotowsky e.a. (2015) demonstrated that just three encounters can already have a significant effect. But even though the study by Zlotowsky e.a. provides a lot of novel insights on the effect of repeated encounters on the uncanny valley phenomenon, it is still the first research into this specific topic. Therefore, several gaps can be identified based on their outcomes.

(11)

11 examples. By adding insights about robots that differ slightly in appearance, the effect of human likeness on likability over several encounters can lead to a broader view of the phenomenon. Also, this could clear up if the differences between different kinds of robots remain relatively constant, which was found for the android and the machinelike robot, or if this effect is only applicable to these specific embodiment types.

Secondly, this study focused on robot perception only, even though human beings are defined to be the best possible option according to the uncanny valley (Mori, 2012). By adding a benchmark to the experiment in the form of a human service provider, the effects of robot encounters can be compared to how perception of humans changes over time. The addition of a human can help clarify further how different levels of human likeness are perceived, as a human being is presented as the most likable option in the uncanny valley theory. It would be interesting to see whether or not habituation effects play a role in repeated encounters with human beings and if this effect manifests in a similar way as has been found for repeated robot encounters.

Thirdly, the perception of human beings over time can provide a framework for interpreting the result of encountering robots over time. It could clarify whether the gap between perceiving human beings and robots stays relatively constant over time or if this could also be subject to a decrease as a result of repeated encounters. Zlotowsky e.a. (2015) showed that the perceived level of eeriness decreases at a somewhat similar rate for different kind of robots. However, the question remains if getting familiar with human beings happens along the same lines and how the gap between human beings and different kinds of robots acts over time.

By building upon these findings and filling these gaps, a more complete picture of the phenomenon can be drawn.

3. Conceptual framework

Based on the given literature review, an expectation of how repeated robot encounters affect outcomes in a service context is formed. Existing literature provides a lot of information about how different types of robots are perceived. However, this paper aims to provide more insight in the processes that occur when a person encounters a robot repeatedly and how these processes affect service outcomes. The proposed conceptual framework is presented in Figure 2. This model will be examined based on the hypotheses that will be specified in the following paragraphs.

Figure 2: Conceptual model – How repeated robot encounters are expected to affect service outcomes The aim of the presented conceptual model is to pinpoint what level of eeriness different kinds of robots evoke in humans during a service interaction. This paper will focus on two key service outcomes which are compliance and loyalty. Compliance is defined as the extent to which customers follow orders of a service provider (Hausman, 2004). Loyalty which is defined as a feeling of allegiance as a result of a service interaction which can be measured through either the intention to repurchase

(12)

12 (Dong, 2011) or the intention to recommend (Ying and Meng, 2009). Compliance in this case is a service aspect that happens during the service interaction, while loyalty can be viewed as the product of a service interaction after this interaction has been finished.

To be able to distinguish between different kinds of robot embodiment, three different kinds of service assistants have been included in the model. As Mori’s initial uncanny valley theory (2012) proposed a variation in levels of evoked eeriness based on higher or lower levels of human resemblance, a similar distinction is made in this model between a machinelike robot and a humanlike robot. Furthermore, a human is added to the mix of service providers to serve as a benchmark to compare the outcomes of both robots to. Also in line with Mori’s theory (2012), eeriness has been defined as a key variable within the model, because according to his study all robots evoke a specific level of eeriness in humans irrespective of their embodiment. However, as this eerie feeling is attributed to the robot’s embodiment and appearance, this effect is an integral part of encountering robots. Therefore, possible outcomes of these encounters have to be a result of the level of eeriness evoked. This results in an expected mediating role for the variable eeriness. Eeriness is in this case defined as the extent to which respondents feel uncanny upon encountering a robot (Mori, 2012).

Finally, as the overall goal of this paper is to add to the current knowledge about robot perception over time, habituation plays a large part in the model. Based on the research by Zlotwosky e.a. (2015), it is expected that the level of eeriness that a robot evokes will decrease after more than one encounter, irrespective of both its embodiment as well as the initial level of eeriness they evoked.

The model and it’s components that have just been described, will be discussed more in depth in the following paragraphs.

3.1 Robots and eeriness

When it comes to the first two gaps in existing literature, it can be concluded that some more insight should be provided by researching the levels of eeriness that different kinds of embodied robots evoke in comparison to human beings, because only the effects a handful of embodied robots have been researched thus far. By adding insights into both different kinds of embodied robots as well as how their impact relates to impact made by human beings, the uncanny valley theory can be enriched to be able to give more specific advise as to what kinds of robots would minimize the risk of evoking an uncanny feeling.

The main theory that forms the basis of the conceptual model is the uncanny valley effect. Based on extensive research, the conclusion can be drawn that moderate to high human resemblance in robots combined with subtle imperfect aspects leads to a sensation of eeriness (Burleigh and Schoenherr, 2015; Mori, 2012; Stein and Ohler, 2016; Zlotowsky e.a., 2015). Even though the larger part of this relationship is linear where an increase in human likeness leads to an increase in positive affect, robots that fall in the valley evoke strong negative emotions. Furthermore, the uncanny valley theory identifies a human being as the top option when it comes to evoking positive affect (Mori, 2012). Therefore, a human being could be used as a benchmark, as it is always expected to resemble the best possible option and thus the best outcomes regarding evoking both the highest levels of affinity and the lowest levels of eeriness. The machinelike robot would come in second place as, even though it evokes lower levels of affinity than a human being, it is expected to perform better than a humanoid robot. As explained by Mori (2012), a robot that closely resembles a human that still possesses imperfections induces the highest levels of eeriness, which leads to expect that the humanlike robot would come in third place when it comes to the level of perceived eeriness evoked.

(13)

13 research has also shown contradicting results. For example, robots that more closely resemble humans have been found to inspire more trust, are perceived more sociable and are more likely to encourage bonding (Van Doorn e.a., 2016). These outcomes suggest a linear effect of human likeness on affinity (Burleigh e.a., 2013). A human being is in this case still expected to evoke the highest levels of affinity, but contrary to the uncanny valley theory, the amount of affinity evoked by a robot increases as the robot becomes a closer reflection of a human being. This would mean that a humanoid robot does not evoke an eerie sensation, but on the contrary would lead to high levels of affinity. This would result in a different ranking than proposed based on the uncanny valley. The top position would remain the same, as a human being is still acknowledged to evoke the highest levels of affinity and thus the lowest levels of eeriness. But the humanoid robot would come in second place and the machinelike robot would come in last place.

Combining these theories, there are similar expectations for both robots that do not resemble humans at all and for actual human beings. Both the uncanny valley theory (Mori, 2012) and the linear effect theory (Burleigh e.a., 2013) agree that machinelike robots evoke low levels of affinity and that human beings evoke the highest levels of affinity. Therefore, it is first of all expected that humans evoke the lowest levels of eeriness as both theories acknowledge a human being as being the best option for evoking positive affect. However, regarding the different kinds of embodied robots, some contradicting findings have been presented in existing literature. The uncanny valley theory proposes very high levels of eeriness and low levels of affinity (Burleigh and Schoenherr, 2015; Mori, 2012; Stein and Ohler, 2016; Zlotowsky e.a., 2015), while the linear effect proposes relatively high levels of affinity in comparison to more machinelike robots (Burleigh e.a.,2013). However, even though human likeness in general increases perceived affinity towards the robot, it is expected that robots with moderate levels of human likeness do induce a sensation of eeriness as presented in the uncanny valley theory. As the moderate levels of human likeness represents the highest levels of eeriness in the uncanny valley theory, the machinelike robot is expected to evoke higher levels of affinity and consequently lower levels of eeriness than the humanoid robot.

Based on this reasoning, the first hypothesis is formulated below. Exploring the effects of these different kinds of embodiments are expected to provide more specific guidelines for designing robots for service settings

H1a) Robots evoke higher levels of eeriness in a service context than humans

H1b) Humanoid robots evoke higher levels of eeriness in a service context than machinelike robots 3.2 Robots and habituation

For the first hypothesis, the focus lies on the perception of a robot during a first encounter. The following paragraphs, however, aim to identify how these perceptions change over repeated encounters. This brings us to the next literature gap related to clarifying this process for robot encounters in comparison to repeated encounters with humans.

(14)

14 that the corresponding levels of eeriness for all kinds of robots decrease over time. Furthermore, these decreases in eeriness are expected to be of similar size, which means that the differences in perception between different kinds of robots stay relatively constant.

If this assumption holds for all different kinds of embodiment and not only the two kinds of robots used by Zlotwosky (2015), it could be expected that people can get used to all kinds of embodied robots. Therefore, based on the findings of Zlotowsky e.a. (2015), it is expected that the gaps in perceived eeriness between the machinelike robot and the humanlike robot will stay relatively constant over time.

Figure 3: Proposed effect of repeated encounters on the uncanny valley (Zlotwosky e.a., 2015). There are several theories are present that can guide the expectations of this phenomenon. First of all, the initial extreme perception of a robot may weaken over time as a result of the affective habituation process (Dijksterhuis and Smith, 2002; Neal e.a., 2009). Second, this initial perception might even turn into a positive feeling as a result of the mere exposure effect (Bornstein, 1989; Burleigh and Schoenherr, 2015; Zajonc, 1968). This leads to expect that even humanlike robots that initially evoke a sensation of eeriness, could over time be experienced in a more positive way. A third theory assigns the initial negative perception of a robot to the novelty of the stimulus. Over time, this feeling could diminish after it has become clear that the stimulus is not a threat (Lee, 2001), leading to a reduction of the negative feeling evoked by the first encounter.

All three of these theories have proven to be at play when it comes to encountering robots specifically (Zlotowsky, 2015). This research showed that the levels of perceived eeriness reduced over time for different kinds of robots. They found a reduction in perceived eeriness of an android robot that could be explained by the affective habituation effect and an increasingly favourable attitude towards a machinelike robot that was explained by the mere exposure effezct. A third explanation provided was the increasing familiarity of an at first sight novel stimulus, leading to easier processing after several encounters with the stimulus (Bornstein and D’Agostino, 1994).

(15)

15 repeated robot encounters and repeated human encounters. This leads to the formulation of the second hypothesis.

H2a) The level of perceived eeriness will decrease over repeated encounters for both machinelike robots and humanoid robots

H2b) The gaps in perceived eeriness between machinelike robots and humanoid robots will stay relatively constant

3.3 Robots and service outcomes

The first two hypotheses treated the different levels of eeriness that can be evoked by different kinds of embodied robots and human beings, at one moment in time as well as over repeated encounters. However, evoking eeriness is not the goal of inserting embodied robots in service situations. Ideally, the robots would result in positive service outcomes, even though they might induce an eerie sensation at first sight. Several scholars have found that higher levels of human resemblance in robots can positively influence service outcomes (Van Doorn e.a., 2017; Broadbent e.a., 2008). As mentioned before, this study will focus on two service outcomes; compliance and loyalty.

The effect of eeriness of compliance

Human beings are known to be able to induce compliance in two ways, which are persuasion and behavioural influence techniques. The first technique relies on influencing the cognitive base on a customer’s behaviour and the second technique relies on influencing behaviour directly (Reingen, 1978). Furthermore, whether or not a customer complies with a request, is known to be influenced by numerous factors like trust in the service provider, a desire to please the service provider, as well as external influences (Hausman, 2004). Other scholars found that the ability to induce compliance can be increased by providing support and encouragement (Vale e.a., 2002) and that compliance can be made more attractive for customers as a result of repeated interaction, reduced effort, education and compensation (Hausman, 2004).

According to Menon e.a. (2004), positive affect towards a service exchange can cause an increase in likelihood to comply. However, compliance is also known to be phenomenon that evolves over time through repeated interactions with a service provider. When a customer becomes familiar with the service provider, a fondness can be developed which increases their trust in the service provider as well as the desire to please the provider by complying with a request. Therefore, when repeatedly confronted with the same service provider, this can cause an increase in positive evaluation (Hausman, 2004). However, a negative emotion like eeriness could be expected to lead to lower levels of compliance as it is not likely to evoke the necessary positive affect towards the service encounter. When there is no sense of trust or wish to please the service provider, customers become less likely to comply.

(16)

16 Based on these insights, it is expected that in service contexts specifically, consumers always strive for compliance to bring the service interaction to fruition for their own benefit, even though eeriness is expected to lead to a decrease in compliance in most other situations. This leads to the formulation of hypothesis 3a, which states that higher levels of perceived eeriness result in higher levels of compliance.

Loyalty

Customer loyalty is a service outcome that can be determined based on two key determinants. First, there is the intention to repurchase an item and consequently displaying loyalty towards a product or service (Dong, 2011). The second determinant is a customer’s intention to recommend a product or service (Ying and Meng, 2009). Loyalty has been known to be a dynamic process that can be subject to delay effects (Chang e.a., 2013), which results in different loyalty outcomes when measured at different points in time after a service encounter. In this case, a distinction can be made between immediate evaluations of a service exchange and a delayed evaluation which depends on a consumers memory and consequently on a customer’s emotions, as emotions are directly connected to the memory of previous experiences. Research suggests the existence of recall bias, as it has been observed that highly arousing emotions increase the likelihood of recalling the event at a later time, for negative as well as positive emotions (McGaugh, 2004). An example of a positive and highly arousing emotion is happiness. An example of a negative and highly arousing emotion is eeriness (Chang e.a., 2013). Furthermore, strong emotions lead to even stronger evaluations over time as people overestimate their high-arousal emotions (McGaugh, 2004). Overall, scholars have agreed that positive emotions lead to higher levels of loyalty (Chang e.a., 2013).

For loyalty, it is known that it can be subject to emotion. Positive emotions can lead to an increase in customer loyalty and negative emotions can lead to a decrease (Chang e.a., 2013). As eeriness can be classified as a negative emotion, this would imply that higher levels of perceived eeriness would result in lower levels of customer loyalty. Furthermore, according to previous research, experiences with a strong emotional effect are often overestimated when subject to delay (Chang e.a., 2013). This would mean that a strong negative first perception of a robot might become more negative during following encounters, which could be explained by the primacy effect theory (Ebbinghaus, 1913). This theory describes the cognitive bias of people recalling a first encounter or primary information better than later encounters or information that was presented later on. Based on these insights combined, the perceived levels of eeriness are expected to decrease over time, which implies that loyalty outcomes will decrease along with eeriness and could even become more negative over time due to overestimation effects of the negative first perception. This leads to the formulation of hypothesis 3b, which can be found below.

H3a) Higher levels of perceived eeriness result in higher levels of compliance H3b) Higher levels of perceived eeriness result in lower levels of loyalty 3.4 Eeriness mediation

(17)

17 within service contexts (Menon et al. 2004). When it comes to loyalty, positive emotions can lead to an increase in customer loyalty (Chang e.a., 2013).

As proposed by the uncanny valley, each kind of robot embodiment inherently induces a certain feeling; either a certain level of eeriness for moderately humanlike robots or a certain level of affection for machinelike robots and humans (Burleigh and Schoenherr, 2015; Mori, 2012; Stein and Ohler, 2016; Zlotowsky e.a., 2015). As the influence of the robots appearance is inevitable during a robot encounter, the level of eeriness that a robot evokes is expected to mediate the effect that different kind of service providers have on the service outcomes compliance and loyalty. This leads to the formulation of the fourth and final hypothesis, which is presented below.

H4) The effect that robots have on the service outcomes compliance and loyalty is mediated by the level of eeriness evoked

4. Research design

To test the hypotheses, a within-person experiment was performed based on a survey. The following section presents the stimuli, procedure, measures, sample and models of the study at hand.

4.1 Stimuli

Three different research assistants were presented to the respondents as stimuli. The three research assistant conditions were the following; a human research assistant, a machinelike robot and a humanlike robot. The three research assistants are displayed below in Figure 4. The condition for each participant was defined beforehand. This resulted in three groups of participants that each encountered only one of the three assistants.

Figure 4: The three research assistants; human, humanlike robot and machinelike robot

4.2 Procedure

The experiment was carried out among students of the University of Groningen. All participants were assigned to one of the three research assistant conditions beforehand. This meant each participant was confronted with only one of the assistants during the experiment.

(18)

18 in an uninterrupted sequence. The reasoning behind this setup was that when interrupted with unrelated information, respondents would be more likely to perceive each different part of the experiment as a new encounter, which in turn leads to more valid and reliable answers for each encounter.

4.3 Measures

To measure eeriness, participants were presented with eight statements related to the level of eeriness they perceived when being confronted with the research assistant. These included whether or not the research assistant was awkward, eerie, creepy, uncomfortable, relaxing, comforting, unnatural and weird. The respondents were asked to rate their agreement or disagreement with the statements on a seven-point Likert scale. All questions were formulated negatively with the exception of two questions asking whether or not the research assistant possessed the positive attributes ‘relaxing’ and ‘comforting’. The scales of these two questions were reversed to make sure the responses were comparable to the responses to the other questions. The overall Cronbach’s alpha for the eight statements was 0,90. To transform the eight statements into one variable indicating each person’s eeriness, the mean value of all questions was taken as a representation of each participants level of perceived eeriness. The average perceived level of eeriness for the total data set was 4.64 on a seven-point scale ranging from completely disagree at 1 to completely agree at 7 (SD=1.40).

To measure compliance, the amount of words used in each writing task was counted. This resulted in a total word count for each of the four encounters and four corresponding tasks. The average amount of words used across all encounters for all participants was 182.60 words per task (SD=122.59). To measure loyalty, participants were asked how much they agreed or disagreed with six statements regarding their opinion about the research lab. These questions were all related to how the participants perceived the research lab throughout the experiment, their intention to recommend the research lab to others and their intention to come back to the lab in the future. The agreement or disagreement was again measured on a seven-point Likert scale. These six statements showed a Cronbach’s alpha of 0,88. Again, the mean value of the statements was taken as the overall level of intended loyalty for each participant. The average level of loyalty across the total data set was 5.43 on a scale of 1 to 7 (SD=1.09).

4.4 Sample

The experiment was carried out among 147 students of the University of Groningen. Out of the in total 147 respondents, 137 managed to complete all sections of the experiment. Therefore, all results were based on the response of 137 participants.

Of all participants, 50 people were confronted with the human research assistant, 48 respondents were confronted with the humanoid robot and the remaining 39 respondents were confronted with the machinelike robot.

Of the in total 137 participants, there were 72 females (52,6%) and 65 men (47,4%). The mean age of the overall group was 22,1 years old. Within the research assistant condition groups, demographics were fairly similar. Within the human condition 38% was male and 62% female with an average of 12,72 years old. For the humanoid robot condition, 54,2% of the participants was male and 45,8% female. The group displayed an average age of 22,9 years old. The final group with the machinelike robot condition included mostly men (52,6%) and a few less women (47,4%). The average age within this condition was 21,7 years old.

4.5 Models

(19)

19 evolutionary, complete, adaptive and robust. To account for the criteria, the model started off with a fairly easy structure and with not that many variables for it to remain simple and leave room for evolution and adaptation. As a model should both be complete and simple, this leads to a trade-off. In this case the simplicity of the model was chosen to be emphasized over the models completeness as parsimonious models are often considered as the most successful (Diebold, 2004). Therefore, a more simple model with less parameters was preferred. Furthermore, it was taken into account that the model should at all times lead to sensible outcomes to account for robustness. Both initial models are displayed below.

𝐶𝑜𝑚𝑝𝑙𝑖𝑎𝑛𝑐𝑒𝑖𝑡 = 𝛼 + 𝛽1𝑅𝑒𝑠𝑒𝑎𝑟𝑐ℎ 𝐴𝑠𝑠𝑖𝑠𝑡𝑎𝑛𝑡𝑖+ 𝛽2𝐸𝑒𝑟𝑖𝑛𝑒𝑠𝑠𝑖𝑡 + 𝜀𝑖𝑡

𝐿𝑜𝑦𝑎𝑙𝑡𝑦𝑖𝑡 = 𝛼 + 𝛽1𝑅𝑒𝑠𝑒𝑎𝑟𝑐ℎ 𝐴𝑠𝑠𝑖𝑠𝑡𝑎𝑛𝑡𝑖+ 𝛽2𝐸𝑒𝑟𝑖𝑛𝑒𝑠𝑠𝑖𝑡 + 𝜀𝑖𝑡

In both models, the ‘i’ is a subscript for each participant and the ‘t’ is a subscript for time. In this case, time is determined as one of the four encounters with the research assistant and thus ranges from one to four. As there were 137 participants who completed all aspects of the experiment, ‘i’ ranges from one to 137. The inclusion of a subscript ‘i' indicates that these variables are expected to be different for each participant and the inclusion of a subscript ‘t’ indicates an expected difference for the variables for each of the four encounters. The beta’s (𝛽) are unknown effect parameters that indicate the change in either compliance or loyalty if the corresponding variable changes. Furthermore, there are an unknown constant (α) that indicates the intercept of the model and (𝜀) that indicates the error term which captures the variation in the dependent variable that is considered random.

For the models functional form, a linear additive model was chosen as there were no interaction effects present that had to be taken into account. For a model that is linear in both parameters and variables, it is assumed that the joint effect of the predicting variables equals the sum of their own effects. Furthermore, the model is fully pooled model which can be concluded from the absence of a subscript for the intercept and the absence of a subscript for the effect parameters. A fully pooled model assumes parameter homogeneity, which means that both the constant (𝛼𝑖𝑡 = 𝛼 for all i) and the effect

parameters are equal for each entity (𝛽𝑖𝑡 = 𝛽 for all i). The result is a standard linear model pooling

all data across both i and t.

5. Results

The following section presents the analyses that were performed to test all hypotheses and the results of these tests. All outcomes are then linked to the hypotheses that were specified based on the conceptual framework.

5.1 Outliers

To check for outliers, all numerical variables were plotted in boxplots. Only one outlier was found for the variable eeriness in the machinelike condition. This outlier displayed a value of 1.750 for eeriness even though the mean level of eeriness in this condition was 5.271. Therefore, the outlier was compared to a range of three standard deviations around the average value for this condition. As the standard deviation was 0.909, the range within values can be considered as not being outliers was between 2.545 and 7.997. As the outlier lies outside of this range, it was assumed to be a significant outlier and the participant corresponding to this outcome was removed from the dataset.

5.2 Robots and eeriness

(20)

20 robot and the machinelike robot. The next question was whether or not humanoid robots evoke higher levels of eeriness than machinelike robots.

To test both components of the first hypothesis, an ANOVA was performed to check for significant differences between levels of eeriness for each research assistant. This test was performed over the whole data set, which means the different encounters were not taken into account. The ANOVA showed a significant effect of the different research assistant conditions on the levels of perceived eeriness (F(2,545)=145.85, p<-0.001). When comparing the human condition to both robot conditions, it was shown that the mean levels of eeriness for the human condition were significantly lower (M=3.56, SD=1.13) in comparison to both the humanoid robot condition (M=5.27, SD=1.27) and the machinelike robot condition (M=5.26, SD=0.93). The results of can be found in Appendix 1.

The boxplot in Figure 5 displays the differences in perceived eeriness for all three conditions.

Figure 5: Boxplot of levels of eeriness per research assistant

(21)

21 Figure 6: Boxplot of levels of perceived eeriness per research assistant for each encounter

Several conclusions could be drawn from these results. First of all, it showed that human research assistants evoke lower mean levels of perceived eeriness than both the humanoid robot as well as the machinelike robot. This was true for the total data set as well as for each of the four different encounters. This leads to the acceptation of hypothesis 1a, stating that robots evoke higher levels of eeriness in a service context than humans

Furthermore, conclusions can be drawn about the different levels of perceived eeriness between the two kinds of robots. When examining the results overall, it was shown that the humanoid robots evoked higher mean levels of perceived eeriness than machinelike robots, even though the difference in mean values was very small. However, when comparing the differences between the two robots for each separate encounter, it was found that for all encounters with exception of the first one, humanoid robots always evoked higher levels of eeriness than machinelike robots. Based on these findings, hypothesis 1b was accepted, which stated that humanoid robots evoke higher levels of eeriness in a service context than machinelike robots.

5.3 Robots and habituation

The following hypotheses concern the differences in levels of eeriness over time. However, as it is assumed that encountering humans more often leads only to higher levels of affinity and not to any levels of eeriness, the human condition were not taken into account for these hypotheses.

The first question to answer was whether or not the perceived eeriness for both robot conditions decreased over time. To test this assumption, an ANOVA was performed for both conditions.

For the humanoid robot, the ANOVA was significant (F(3,188)=4.16, p=0,007) and displayed significant differences in levels of perceived eeriness between the four encounters. When compared to the first encounter, all other encounters displayed significant increases in mean levels of perceived eeriness. The first encounter showed the lowest level of perceived eeriness (M=4.75, SD=1.20), followed by higher levels of eeriness in the second encounter (M=5.31, SD=1.11) and an even higher mean value in the third encounter (M=5.57, SD=1.29). For the fourth encounter however, the eeriness decreased slightly (M=5.47, SD=1.33).

(22)

22 Lastly, the test was performed for the human condition, which was insignificant (F(3,196)=0.08, p=0.97). This implied that, just like expected, encountering a human more times does not lead to significant differences in the level of perceived eeriness. The mean level of eeriness in the first encounter was 3.53 (SD=0.94), followed by a mean value of 3.60 (SD=1.15) for the second encounter, a mean value of 3.51 (SD=1.15) for the third encounter and 3.59 (SD=1.29) for the fourth and final encounter.

The differences between the encounters for all conditions can be found in the boxplot in Figure 7. The results of the ANOVA’s can be found in Appendices 6-8.

Figure 7: Boxplot of levels of eeriness at four encounters for the humanoid and machinelike robots Based on these findings, it can be concluded that only the humanoid robot condition displayed significant differences in mean levels of eeriness for all encounters. However, these differences were not in line with expectations. It was expected that for this robot condition the perceived level of eeriness would decrease. However, the perceived eeriness increased in both the second and the third encounter. Only for the fourth encounter, a slight decrease in eeriness in comparison to the previous encounter was found. As the machinelike robot condition did not display any significant difference, the hypothesis could not be accepted, as it stated that the level of perceived eeriness will decrease over repeated encounters for both machinelike robots and humanoid robots.

Gap between humanoid and machinelike robots over time

Building upon the changes in eeriness over time for both robots, the question remained if the gaps between the two robots remained relatively constant over time. To test this hypothesis, again an ANOVA test was performed to check if there were significant differences in mean levels of eeriness between both conditions.

(23)

23 stated that the gaps in perceived eeriness between machinelike robots and humanoid robots will stay relatively constant over time.

5.4 Robots and service outcomes

To be able to test the following hypotheses, two models were specified to calculate the service outcomes compliance and loyalty. These hypotheses concern whether or not changes in the levels of perceived eeriness lead to different levels of compliance and loyalty. To recap, the assumptions are that higher levels of eeriness result in higher levels of compliance and that higher levels of eeriness result in lower levels of loyalty.

Both the initial models for compliance and loyalty that were presented earlier, were basic regression models. However, the data at hand is panel data, which means that for each person there are four encounters resulting in four rows of corresponding levels of eeriness, compliance and loyalty per respondent. First of all, it was necessary to control for the different encounters for each participant within both models. There are several ways to include time in a regression model for panel data. One possibility is to include dummies for all possible times, which results in an encounter specific effect for all units. This approach is appropriate when effects for each specific encounter are expected. Another possibility is to include a time trend. This would result in an increasing or decreasing effect with equal steps for each next encounter. In this case however, the interest lies with encounter-specific changes instead of trends over time. Therefore, four dummy variables were included to take the four different encounters into account. This resulted in the following two models, where T1, T2, T3 and T4 represent the time dummies.

𝐶𝑜𝑚𝑝𝑙𝑖𝑎𝑛𝑐𝑒𝑖𝑡 = 𝛼 + 𝛽1𝑅𝑒𝑠𝑒𝑎𝑟𝑐ℎ 𝐴𝑠𝑠𝑖𝑠𝑡𝑎𝑛𝑡𝑖+ 𝛽2𝐸𝑒𝑟𝑖𝑛𝑒𝑠𝑠𝑖𝑡+ 𝛽3𝑇2 + 𝛽4𝑇3 + 𝛽5𝑇4 + 𝜀𝑖𝑡

𝐿𝑜𝑦𝑎𝑙𝑡𝑦𝑖𝑡 = 𝛼 + 𝛽1𝑅𝑒𝑠𝑒𝑎𝑟𝑐ℎ 𝐴𝑠𝑠𝑖𝑠𝑡𝑎𝑛𝑡𝑖+ 𝛽2𝐸𝑒𝑟𝑖𝑛𝑒𝑠𝑠𝑖𝑡+ 𝛽3𝑇2 + 𝛽4𝑇3 + 𝛽5𝑇4 + 𝜀𝑖𝑡

Pooling versus unit-by-unit models

The current models were both pooled models, which means that all entities are considered to be representable by one regression line. In this case, it would mean one regression line that represents all four encounters and the same coefficients applying across encounters. On the one hand, a fully pooled model offers efficiency benefits and can therefore be convenient to use. A unit-by-unit model on the other hand, provides maximum flexibility for potential differences between entities (Leeflang e.a., 2015). Unit-by-units models require a model estimates for each entity, in this case each encounter, which would result in different parameter estimates for each independent variable. This would result in the following models. For both models, an index t is included for all beta’s and the intercepts, which means that a different beta and a different intercept applies for each encounter t. This resulted in the following unit-by-unit models.

𝐶𝑜𝑚𝑝𝑙𝑖𝑎𝑛𝑐𝑒𝑖𝑡 = 𝛼𝑡+ 𝛽𝑡1𝑅𝑒𝑠𝑒𝑎𝑟𝑐ℎ 𝐴𝑠𝑠𝑖𝑠𝑡𝑎𝑛𝑡𝑖+ 𝛽𝑡2𝐸𝑒𝑟𝑖𝑛𝑒𝑠𝑠𝑖𝑡+ 𝛽3𝑇2 + 𝛽4𝑇3 + 𝛽5𝑇4 + 𝜀𝑖𝑡

𝐿𝑜𝑦𝑎𝑙𝑡𝑦𝑖𝑡 = 𝛼𝑡 + 𝛽𝑡1𝑅𝑒𝑠𝑒𝑎𝑟𝑐ℎ 𝐴𝑠𝑠𝑖𝑠𝑡𝑎𝑛𝑡𝑖+ 𝛽𝑡2𝐸𝑒𝑟𝑖𝑛𝑒𝑠𝑠𝑖𝑡+ 𝛽3𝑇2 + 𝛽4𝑇3 + 𝛽5𝑇4 + 𝜀𝑖𝑡

(24)

24 the degrees of freedom of the by-unit model. V2 stands for the degrees of freedom of the unit-by-unit model.

First, it was tested if pooling was allowed for the Compliance model. The RSS of the pooled model was 7,058,300 with 537 degrees of freedom, while the RSS of the unit-by-unit model was 7,004,500 with 528 degrees of freedom. This resulted in an insignificant F of 0.451 (p=0.907). This means that all parameters did not significantly differ over time, from which it could be concluded that pooling the data was allowed for the compliance model. The results from both the pooled and the unit-by-unit model can be found in Appendices 13 and 14.

Second, the Chow test was performed for the Loyalty model. The RSS for the pooled model was 540.42 with 537 degrees of freedom. The RSS for the unit-by-unit model was 476.40 with 528 degrees of freedom. The resulting significant F statistic was 7.884 (p<0.001). This means that all parameters were significantly different from each at each encounter. From this it could be concluded that pooling was not allowed for the loyalty model. The results from both the pooled and the unit-by-unit model can be found in Appendices 15 and 16.

Fixed time effects

Previously, fixed time effects were included in both models in the form of time dummies. For time fixed effects, it is assumed that there is a correlation between the entity’s error term, in this case the encounter, and the independent variables. However, a test should be performed to check whether or not fixed time effects are actually suitable for the data at hand. To perform this check, both a model with and a model without the fixed time effects were estimated. These models were then compared using two tests which were an F-test for individual effects and a Breusch-Pagan time effects test. This checks if the dummies for all encounters are equal to 0 and thus if the model without fixed time effects showed to fit the data better than the model including fixed time effects.

First, the check was performed for the Compliance model. The F-test for individual effects was performed which resulted in a significant F-statistic of 51.593 (p<0.001). Second, a Breusch-Pagan time effects test was performed, which resulted in a significant chi-square value of 497.47 (p<0.001). Both tests lead to the conclusion that significant fixed time effects were present for the compliance model. For the Loyalty model, the procedure was repeated. This resulted in a significant F-statistic of 48.511 (p<0.001) for the F-test for individual effects and a significant chi-square value of 198.83 (p<0.001) fo r the Breusch-Pagan time effects test. Again, the conclusion was drawn that significant fixed time effe cts were present.

However, as the Chow test for the Compliance model showed that this data was best represented by a pooled model instead of a unit-by-unit model, no further tests were performed to examine either fixed or random time effects. The conclusion was drawn that the best fit for the Compliance data was a pooled model with fixed time effects in the form of time dummies.

Fixed versus random effects

(25)

25 over time. Therefore, both models with random effects would look as follows, where 𝜀𝑖𝑡 represents

the within-entity error and µ𝑖𝑡 represents the between-entity error and without the time dummies

that represented fixed effects.

𝐿𝑜𝑦𝑎𝑙𝑡𝑦𝑖𝑡 = 𝛼𝑡 + 𝛽𝑡1𝑅𝑒𝑠𝑒𝑎𝑟𝑐ℎ 𝐴𝑠𝑠𝑖𝑠𝑡𝑎𝑛𝑡𝑖+ 𝛽𝑡2𝐸𝑒𝑟𝑖𝑛𝑒𝑠𝑠𝑖𝑡+ µ𝑖𝑡+ 𝜀𝑖𝑡

For the Loyalty model, both the fixed effects model and the random effects model were estimated. The fixed time effects model displayed a significant F-statistic of 17.20 (p<0.001) as well as the random time effects model which resulted in an F-statistic of 17.23 (p<0.001). This means that for both models all coefficients were significantly different from zero. To test which of the model represented the data best, a Hausman test was performed. This would test whether or not the unique errors were correlated with the regressors. This test was not significant (ChiSq=0.004, p=1.00), which lead to the conclusion that the random effects fit the Loyalty model best. This thus resulted in a unit-by-unit model with random time effects. The results of the fixed effects model, the random effects model and the Hausman test can be found in Appendices 17-19.

Interaction between time and eeriness

Another way to account for the role times plays in panel data, is to include time as an interaction effect. In this case, this would result in interactions between the independent variable eeriness and the different encounters. To check whether or not this approach leads to a better fit for either Compliance or Loyalty, two new models were specified. This resulted in the following models.

𝐶𝑜𝑚𝑝𝑙𝑖𝑎𝑛𝑐𝑒𝑖𝑡 = 𝛼 + 𝛽1𝑅𝑒𝑠𝑒𝑎𝑟𝑐ℎ 𝐴𝑠𝑠𝑖𝑠𝑡𝑎𝑛𝑡𝑖+ 𝛽2𝐸𝑒𝑟𝑖𝑛𝑒𝑠𝑠𝑖𝑡+ 𝛽3𝐸𝑒𝑟𝑖𝑛𝑒𝑠𝑠𝑖𝑡∗ 𝑇2

+ 𝛽4𝐸𝑒𝑟𝑖𝑛𝑒𝑠𝑠𝑖𝑡∗ 𝑇3 + 𝛽5𝐸𝑒𝑟𝑖𝑛𝑒𝑠𝑠𝑖𝑡∗ 𝑇4 + 𝜀𝑖𝑡

𝐿𝑜𝑦𝑎𝑙𝑡𝑦𝑖𝑡 = 𝛼 + 𝛽1𝑅𝑒𝑠𝑒𝑎𝑟𝑐ℎ 𝐴𝑠𝑠𝑖𝑠𝑡𝑎𝑛𝑡𝑖+ 𝛽2𝐸𝑒𝑟𝑖𝑛𝑒𝑠𝑠𝑖𝑡+ 𝛽3𝐸𝑒𝑟𝑖𝑛𝑒𝑠𝑠𝑖𝑡∗ 𝑇2

+ 𝛽4𝐸𝑒𝑟𝑖𝑛𝑒𝑠𝑠𝑖𝑡∗ 𝑇3 + 𝛽5𝐸𝑒𝑟𝑖𝑛𝑒𝑠𝑠𝑖𝑡∗ 𝑇4 + 𝜀𝑖𝑡

By including the different encounters as interaction effects, the model accounts not only for the effect that an encounter has on the dependent variable, but also accounts for the effect each new encounter has on the level of perceived eeriness. As eeriness is expected to change over time, it can be assumed that an interaction is at play between eeriness and the different encounters.

However, as including the interaction effect leads to quite a few similar variables, this increases chances of multicollinearity. Multicollinearity can be defined as high degree of correlation between the predicting variables which leads to unreliable parameter estimates (Leeflang e.a., 2015).

To test for multicollinearity, the correlation matrix, variance inflation factor (VIF scores) and tolerance were examined. According to Leeflang e.a. (2015), VIF scores should be below 5, tolerance should be higher than 0.2 and the correlation R2_{should be below 0.5. The formula’s for the VIF scores and} tolerance are presented below.

(26)

26 Second, running the new regression model for Loyalty resulted in a relatively low adjusted R2_{of 0.1513.} The F-statistic was significant however (p<0.001) and estimated at 17.13. In this case, both the humanoid and the machinelike condition of the Research Assistant showed significant effects. The humanoid robot was estimated at 0.41 (p<0.001) and the machinelike robot at -0.37 (p=0.003). Furthermore, the interaction between Eeriness and the second encounter was significant (p<0.001) and estimated at -0.13. The results of the interaction model can be found in Appendix 21.

Third, the models were tested for multicollinearity. As both models were identical except for their dependent variables, the tests for multicollinearity were identical for both models.

When examining the correlation matrix of all dependent variables, the correlation between the variables Eeriness and Research Assistant exceeded the R2 _{value of 0.5 (R}2_{=0.52, p<0.001). No other} variables exceeded this threshold of 0.5. All R2 _{values can be found in Appendix 22. However, as both} variables are essential to the model, none of the variables was excluded.

Next, the VIF scores were assessed, which were very high for all variables with the exception of the variable Research Assistant. All other VIF scores exceeded the threshold of 5, which is a sign that collinearity is a problem in the model (Leeflang e.a., 2015). This is only confirmed by the very high scores ranging from 5.799 for Eeriness to 22.183 for the interaction between Eeriness and T2. All scores can be found in Appendix 23.

Finally, the tolerance for each variable was examined. Again, the results showed that the model suffers greatly from collinearity, as all variables except for the Research Assistant showed tolerance levels of below the 0.2 threshold.

Based on the outcomes of examining the correlation matrix, the VIF scores and the tolerance levels, the conclusion was drawn that there is too much collinearity present within the model. According to Leeflang e.a. (2015), no solution to a regression model can be obtained in the case of extreme multicollinearity. Therefore, it was determined that this model was not suitable for both Compliance and Loyalty.

Effects on compliance over time

As mentioned before, a pooled model including time dummies for fixed time effects was found to be the best fit for the Compliance data. This resulted in one set of estimator parameters for the total data set, as the pooled model implied that the data could be pooled over the four different encounters. The model can be found below.

𝐶𝑜𝑚𝑝𝑙𝑖𝑎𝑛𝑐𝑒𝑖𝑡 = 𝛼 + 𝛽1𝑅𝑒𝑠𝑒𝑎𝑟𝑐ℎ 𝐴𝑠𝑠𝑖𝑠𝑡𝑎𝑛𝑡𝑖+ 𝛽2𝐸𝑒𝑟𝑖𝑛𝑒𝑠𝑠𝑖𝑡+ 𝛽3𝑇2 + 𝛽4𝑇3 + 𝛽5𝑇4 + 𝜀𝑖𝑡

The model showed that compliance could be significant explained by the independent variables (F=13.98, p<0.001). The model displayed an adjusted R2_{value of 0.125, which means that the included} variables could account for 12,5% of the variation in the dependent variable. Taking a closer look at the independent variables, only two showed to have a significant effect on compliance. These variables were the time dummy for the third encounter (β=-72.64, p<0.001) and the time dummy for the fourth encounter (β=-105.71, p<0.001). From these results, it could be concluded that the different Research Assistants and perceived levels of eeriness had no significant effect on compliance.

However, the model did show a decrease in compliance over time. As the third encounter showed a significant decrease in eeriness in comparison to the first encounter and the fourth encounter showed an even larger decrease in eeriness, the conclusion could be drawn that compliance decreases over after several encounters irrespective of both the research assistant a participant was confronted and the level of perceive eeriness this participant experienced. The results of the model can be found in Appendix 13.

Can we get used to robots? Exploring the uncanny valley effect over time