Looking at different environments, previous research has shown that the interaction between a hu- man-like robot and children was able to leverage the motivation, sense of community, and self-expres-sion in the education domain (Han, Park & Park 2015). In shopping malls, the interaction between a robot and humans was perceived as enjoyable and benefits from this interaction such as clear instruc-tions were derived (Satake et al. 2015). Acceptance Since 90% of the people in a shopping mall stated that they want to use the robot again, high ac-ceptance of robots was found and 65% even preferred a robot over a human counterpart (Satake et al. 2015). In the medical environment, the acceptance of robots was similarly high. In one study, only 6% of participants did not want the presence of a robot (Yoshikawa et al. 2011). This might be due to advantages of robots like better availability, higher accuracy and speed, and lower distractions com-pared to traditional doctors (Broadbent et al. 2010). However, there was not always a positive effect. Some participants experienced technological mal-functions of the robot and therefore were frustrated about the usage or saw it as an invasion into privacy (Broadbent et al. 2014, Orejana et al. 2015). Other participants simply did not care if they use a robot or not (Robinson, Broadbent & MacDonald 2015). Some people have concerns about the reli-ability and capability of robots when interacting with them (Broadbent et al. 2010). Others rate a too human-looking robot as scary and therefore would avoid interacting with it (Satake et al. 2015). In another previous research in a shopping mall, half of the participants even feared the interaction with a robot and therefore preferred to have a human counterpart (Ogawa et al. 2011). As there are positive and negative reactions towards the interaction with robots, it is important to know how a robot has to look like and how he should behave and react to interactions by humans. Appearance Looking at the appearance of robots, one can see that not only one type of robot is used for example within the domain of elderly care. Previous research found that depending on the job the robot does a more animal-like robot or a more machine-like robot is preferred (Broadbent et al. 2009). For health-related tasks like reminding people of medication a machine-like robot was preferred, whereas for companionship a more animal-like robot was favored. Another study found that since robots should not serve as human substitutes they also do not need to look like humans and machine-like robots were preferred over human-like robots due to the association with functionality (Broadbent et al. 2012).
fact that people have the desire to be viewed by others in a positive way (Leary, Kowalski 1990). This reinforces the effect social presence can have. Other studies found that social presence often results in negative effects. Latané & Harkins (1976) found that an increase in the number of people in an audience leads to higher tension in an individual. Griffit & Veitch (1971) similarly found that the feeling of comfort decreases as more people are present.
The concept of non-interactive social presence is based on the social impact theory (Argo, Dahl & Manchanda 2005). This theory is composed by three principles (Latané 1981). The first principle is the social force. This force is a multiplicative function of the importance, the proximity, and the number of other people. The theory says that the impact of social presence on a person increases as the im- portance of the group, the proximity, and the number of people increases (Latané, Wolf 1981). Com-bining all those three aspects leads to the greatest influence of a group. Theory says that while a present social group is increasing in size it leads to higher negative emotions such as nervousness, tension, and anxiety (Jackson, Latané 1981, Latané, Harkins 1976). While this linear relationship might hold for interactive social presence, theory states that for non-interactive social presence a v-shaped relationship applies (Argo, Dahl & Manchanda 2005). This means that the most positive emotions were caused where only one person is present while more or less people pre-sent leads to more negative emotions. This relationship is due to the fact that people have a motivation to belong and therefore appreciate a small number of persons present (Baumeister, Leary 1995). Regarding the proximity of the non-interactive group, theory says that a close group of a large size of people arouses in ourselves the feeling that our personal space is being threatened and thus negative feelings such as pressure, stress and discomfort increase (Argo, Dahl & Manchanda 2005, Dabbs 1971, Sommer 1969). With respect to the importance of a non-interactive group in form of status, power, and ability (Latané, Wolf 1981), no difference between the influence of a group of friends or strangers could be found (He, Chen & Alden 2012). This is why this factor is not included in this research. The second principle of social impact theory is the psychosocial law (Latané 1981). This principle states that the impact of social presence has a marginally decreasing effect. This means that the effect of the presence of the first person in the surrounding of an individual is greater than the impact of the tenth person. The third principle describes the impact of division. It means that when there are more indi-viduals present, the impact of a social group is divided among the individuals.
Since social impact theory suggests that with a non-interactive social presence negative emotions arise, it can be hypothesized that
3 Methodology In this chapter, firstly the method of conjoint analyses as a measurement of preferences with its un-derlying models is explained. Then it is elaborated on the experimental design and the implementation into a questionnaire. 3.1 Preference measurement In order to measure the preference structure of participants, conjoint analysis is the most popular method (Eggers, Sattler 2011). Conjoint means that respondents evaluate several shown products by considering the different product attributes and their levels jointly. It is a good method for testing innovations and having insights for new product developments (Urban, Hauser & Urban 1993, Sattler 2006). Also, it can be used to identify segments in a population that have different preferences for the attributes (Teichert 2001). These preferences should be homogeneous within the segments but heter-ogeneous between them. Since the different classes are latent, each participant belongs with a certain probability to one of the segments (DeSarbo, Ramaswamy & Cohen 1995). From different existing methods of conjoint analyses like rating-based conjoint, ranking-based con- joint, and choice-based conjoint, the latter method is used for this research. This is because choice- based conjoint represents choices people have to do in everyday life the best and therefore is an ef-fective approach (Eggers, Sattler 2011). Additionally, a no-choice option can be added which increases the realism of the choice task. 3.2 Models Hereafter, the utility model and the choice model which are the underlying models of conjoint analyses are elaborated. 3.2.1 Utility Model
The underlying concept that models the choice of a decision maker when choosing from different product alternatives is the random utility theory (Walker, Ben-Akiva 2002). While a participant is mak-ing an observable choice among the alternatives, an unobservable utility for the chosen alternative is assumed. This utility ! for a product "[" = 1, … , ()] for individual +[+ = 1, … , ,] is composed of a systematic utility term - and a random disturbance term ε which leads to the following equation:
!/)= -/)+ ε/) . (1)
2011). These estimates are part-worth utilities of individual + for attribute 2. This leads to the following equation for the systematic utility - of individual + for product ": -/) = 76895)61/6 . (2) Random utility theory assumes that individuals want to maximize the utility (Walker, Ben-Akiva 2002). That means that individual + chooses product " only if !/)≥ !;) for all < ∈ >) , (3)
Table 2: Attributes and levels.
Attribute Level 1 Level 2 Level 3
Service partner Human-like robot Machine-like robot Human Location At the counter In a separated room ---
Waiting time None 3 minutes 6 minutes
In the second part, participants have to do the choice task. Within a choice set, participants have to indicate their most preferred out of three options and if they would actually use it in real life. Each of the options is a unique combination of different levels of the attributes. In the third part, participants are asked general questions about their demographics and personality. The information about demographics can be provided voluntarily. In the end, contestants also have to state which is the fictitious task they had to do in the bank. This control question is implemented to see if people were paying attention throughout the survey. After that the survey is completed. 4.2 Sample characteristics A total of 241 persons completed the question-naire. Eleven of them did not answer the last control question correctly and therefore were excluded from the sample for the analysis. This leads to a total of 230 participants of the final sample. Specific characteristics of the respond-ents within each scenario can be found in Table 3. Of the final number of participants, 117 were allocated to scenario 1 (easy condition) and 113 to scenario 2 (complex condition).
Combining both scenarios, respondents needed 8:59 minutes to complete the ques-tionnaire. A total of 144 females (62.6%) and 83 males (36.1%) took part in this survey. The average age in the sample is 26.92. Most par- ticipating people of the survey come from Ger-many (n=154, 67.0%). Regarding education level, 98 persons (42.6%) indicated that they have a non-university degree and 131 (57.0%) stated that they have a completed university degree. Of all respondents, 56 persons (24.3%) are students, 62 (27.0%) are employed, and 104 (45.2%) are unemployed. Since respondents are rather young students or unemployed, most peo-ple (n=123, 53.5%) have a yearly income below EUR 15,000. A total of 47 persons (20.4%) indicated that they have a yearly income between EUR 15,000 and EUR 50,000 and 12 (5.2%) stated they have an income of more than EUR 50,000 per year. Table 3: Sample characteristics. Scenario 1 Scenario 2
Variable Sample Sample
To check whether the items of the personality question load on one factor each, factor analyses were conducted. As depicted in Appendix B, the four items of technology readiness and privacy concerns as well as the three items for embarrassment and task perception load on one factor. The outcome of the personality questions can be found in Table 4. Reliability analyses of the scales are conducted. They show that the items of technology readiness, privacy concerns and task perception can be combined in one factor since Cronbach’s Alpha is always higher than 0.6 and the value always decreases when one item is deleted. For the scale of embarrass-ment, Cronbach’s Alpha is also larger than 0.6 but excluding the item of insecurity increases this value even more. Therefore, this scale is split into two parts, one combining ashamedness and embarrass-ment and the other one insecurity. Based on the means, the technology readiness of respondents can be seen as moderate (4.3) and people are severely concerned about the privacy of personal infor- mation (5.2). Regarding the interaction with robots, respondents only feel slightly shamed or embar-rassed (2.8), but are more insecure (4.0). In the first scenario (low complexity), the task is perceived as easy (2.4). In the second scenario (high complexity), the task is not perceived as very hard but more Table 4: Outcome of personality traits.
Mean (Std. dev.) Cronbach's Alpha Factor mean
Technology readiness 0.668 You prefer to use the most advanced technol-ogy available. 4.6 (1.5) 4.3 It seems your friends are learning more about the newest technologies than you are. (Re-coded) 4.0 (1.6) In general, you are among the first in your cir-cle of friends to acquire new technology when it appears. 3.5 (1.5) You can usually figure out new high-tech prod-ucts and services without help from others. 5.3 (1.3) Privacy concerns 0.888 It usually bothers me when companies ask me for personal information. 5.1 (1.4) 5.2 When companies ask me for personal infor- mation, I sometimes think twice before provid-ing it. 5.4 (1.4) It bothers me to give personal information to so many companies. 5.2 (1.5) I'm concerned that companies are collecting too much personal information about me. 5.2 (1.5) Embarrassment 0.817 Ashamed 2.7 (1.5) 2.8 Embarrassed 3.0 (1.7) Insecure 4.0 (1.8) 4.0
Scenario 1 Scenario 2 Scenario 1 Scenario 2
difficult (3.0) than in the contrary scenario. To test whether the difference of task perception in the two scenarios is significant, a one-way ANOVA is run. As the results of the analysis in Table 5 indicate, the difference in task perception is significant (p<0.001).
Table 5: ANOVA table of difference between task perception mean.
Squares df Sum of Square Mean F Sig.
Difference in per-ceived task difficulty Between Groups (Combined) 1757.483 1 1757.483 1048.751 0.000 4.3 Choice-based conjoint analysis In the following, a decision about how to integrate the variable waiting time is done. Then several models are compared and the model fits are assessed. Thereafter, the main effects and moderation effects are analyzed. The predictive validity of the best-fitting model is derived as well as the willing-ness-to-wait. For the analysis of the data of the survey a subset that contains 15 of the 18 choices sets is used. The remaining three choice sets are later used to measure the predictive validity. 4.3.1 Part-worth or linear model In the model specification, the format of the attributes must not be the same for all. In this survey, the attributes “service partner” and “location” are nominal and therefore a part-worth model. However, the attribute “waiting time” is not nominal and therefore can be a part-worth or a linear model. To test which one of the two possibilities is better for including the waiting time, two different models are estimated, one which includes waiting time as a part-worth model and one which includes waiting time as a linear model. The results of these estimations can be found in Table 6. Table 6: Comparison of part-worth and linear waiting time models.
In contrast, Model 4 has a higher log-likelihood, a lower AIC and a slightly higher McFadden adjusted R², which speaks for Model 4. However, since models with a lower number of parameters are preferred and the difference in the adjusted R² is very small, Model 3 is preferred over Model 4 which again indicates that there is no moderating effect of social presence. 4.3.3 Main effects For testing the main effects, Model 1 is used. The results of the parameter estimation can be found in Table 8. Looking at the Wald statistic, the table shows that the differences between the parameter estimates within the attributes are all significant on a 1% significance level. The z-value for each level shows that the estimates for the parameters are all significantly different from zero on a 1% signifi-cance level. The calculated relative importance shows that the service partner (48.6%) is the most important attribute for the participants of the survey. This attribute is closely followed by the waiting time (44.3%). The location (7.1%) has a very low relative importance compared to the other two at-tributes. The estimate for the no-choice-option (β=-2.2828; p<0.01) shows that using the service is preferred by participants over not using any of the offered services. Table 8: Results for main effects.
Attributes Utility z-value Importance Relative Wald p-value
Table 9: Moderating effects of non-interactive social presence on type of service partner.
Effect of location on… Utility z-value Wald p-value
Human-like robot -0.0124 -0.30 0.09 0.76 Machine-like robot 0.0747 1.93** 3.74 0.053 Human -0.0623 -16.05*** 25.74 0.11 ***The P-Value is < 0.01; **the P-Value is < 0.05. Secondly, the moderating effect of the task complexity on the choice of a service partner is investi-gated. In scenario 1 the task complexity is low whereas in scenario 2 the task complexity is high. For analyzing the moderator, Model 3 is used. The results of the estimation can be found in Table 10. The differences between parameter estimates within an attribute are all significant at a 1% significance level. The Wald(=) statistic indicated that the respondents in the two scenarios have different prefer-ences for each attribute. Most of the parameter estimates are significantly different from zero at a 5% significance level, except the location estimates in scenario 1 (p=0.2643) and the three minutes waiting time in both scenarios (p=0.1867 in scenario 1; p=0.2177 in scenario 2). Looking at the relative im- portance, one can see that the preferences of participants in both scenarios differ. Whereas for re-spondents in scenario 1 waiting time (60.4%) is most important, for respondents in scenario 2 the type of service partner (57.9%) is most important. In the low complexity scenario, the type of service part- ner (38.8%) is second important and the location (0.9%) is least important. In the high complexity sce-nario, the waiting time (25.2%) is the second important attribute and also the location (16.9%) is the least important one, though more important than for respondents in scenario 1. Hypothesis 4a assumes that in a complex situation, human-like robots are more preferred over ma-chine-like robots than in an easy situation. Looking at the parameter estimates of human-like robots (β=-0.8634; p<0.01) and machine-like robots (β=-0.0872; p<0.05) in scenario 1 and human-like robots (β=-0.8056; p<0.01) and machine-like robots (β=-0.7005; p<0.01) in scenario 2 respectively, shows that both times a machine-like robot is preferred over a human-like robot. Therefore, H4a is not supported. Table 10: Results of moderating effect of task complexity. Scenario 1 Scenario 2 Attributes
Utility z-value Imp. Rel. Utility z-value Imp. Rel. Wald p-value Wald(=) p-value
11 one can find the number of cases that are predicted correctly as well as the total number of cases for each scenario.
Table 11: Hit rate of Model 3.
this value is larger than the highest value stated in the survey (6 minutes). Therefore, it could be that utilities for people might change after waiting 6 minutes. In the calculation, it was assumed that the trend between no waiting time and waiting 6 minutes continues beyond this value. Figure 2: Probability of preferred service vs. no-choice option. 4.3.7 Hypotheses summary
In Table 13, an overview of the tested hypotheses and the results of this research can be found. Table 13: Overview of hypotheses.
Hypotheses Supported
4.4 Segmentation Hereafter, firstly an appropriate number of segments is determined. In a next step, those segments are described. The segmentation of the sample was done with the data of the 15 choice sets that were also used for the analysis of the hypotheses. The number of segments is unknown prior to the analysis. Therefore, several models for different numbers of segments are estimated and compared among each other. The best-fitting model is chosen based on information criteria and classification errors of each model. 4.4.1 Number of segments To find the optimal number of segments, seven different models with two to eight classes are esti-mated. An overview of the results of these estimations can be found in Appendix C. All of these splits are significant at a 1% significance level. A graphical comparison of the information criteria which are used to find the best model can be found in Figure 3. To determine the optimal number of segments, the elbow criterion is used. This means that the ap-propriate number of classes can be found where there is a bend in the graph (Ketchen Jr, Shook 1996). Figure 3 shows that for Model 7 there is an elbow in all of the graphs of the different information crite-ria. Looking additionally at the classification errors of the different models shows that Model 7 (0.0094) has the second lowest value after Model 5 (0.0078). Since Model 5 has very large values for the infor-mation criteria compared to Model 7, the latter one which has four different segments is chosen for analysis. Figure 3: Information criteria comparison. 3500 3700 3900 4100 4300 4500 4700 4900
Model 5 Model 6 Model 7 Model 8 Model 9 Model 10 Model 11
4.4.2 Interpretation of segments The preferences of the four different segments and the demographic and personality characteristics can be found in Table 15 and Table 14, respectively. Looking at the Wald(=) statistics, one can see that people in each class have significantly different pref- erences for each attribute (p<0.01). Also, the technology readiness (p=0.024), the feeling of embar-rassment (p<0.01) and insecurity (p<0.01) while using a robot, and the lowest income class (p<0.01) are significant covariates to do the segmentation. All segments are willing to choose the service they are offered since their beta-values for the no-choice option are negative (ranging from β=-2.9593 (p<0.01) to β=-0.4262 (p<0.01)).
Class 1 – Insecure old-fashioned Ger-mans
This segment is with 34.3% the big-gest one. It consists mostly of cus- tomers who had to perform the diffi-cult task. It is the sector with the highest privacy concerns and also feels most insecure while interacting with a robot. Compared to the aver-age, there are most Germans in this segment and participants do not have a university degree. The most important attribute for this segment is the kind of service partner (98.5%). Only the parameters for this attrib-ute are significantly different from zero (p<0.01). Most preferred is a hu- man service partner which is congru-ent with the fact that they feel most insecure while interacting with a ro-bot. However, this segment is the one that is most unlikely to choose the no-choice option. For managers, it is very desirable to focus on this
Table 14: Demographics and characteristics of segments.
Class 1 Class 2 Class 3 Class 4
why a machine-like robot is preferred over a human-like robot in this research. However, the most preferred option was the human-being who is most familiar with us. This result is not surprising but it shows that most of the persons are not yet ready to engage with modern technology. To answer the first research question, it means that having either a human-like or a machine-like robot in a bank would influence the choice of consumers in a negative way. The finding that machine-like robots are preferred over human-like robots is also not dependent on the task participants had to do. Research in the past found that human-like robots are perceived as more intelligent (Walters et al. 2008) which suggests that they should be preferred over machine-like robots when the task someone has to do is complicated. However, this study found that there is no difference between the two kinds of robots if the task is complicated. In fact, the human is much more preferred over any kind of robot and it gets also much more important for people to have a human as a service partner compared to when they are doing a less complex task. This can be explained by the finding from previous research that people have doubts about the capabilities of robots (Broadbent et al. 2011). In a more complex task the capabilities of the service partner is of more importance and therefore it is also more important to be served by a human. Still, it has to be acknowledged that there are people that would also use a machine-like robot. This study has shown that there is a segment for which using this kind of robot does not lead to a decrease in the utility of the service. To answer the second research question, it means that the complexity influences the choice of con-sumers to that extent that while engaging in a difficult task any kind of robot is even less preferred over a human. Regarding social presence, a lot of previous research found that the mere presence of other people around us has an influence on our behavior (Argo, Dahl & Manchanda 2005, Dahl, Manchanda & Argo 2001, He, Chen & Alden 2012) and that the embarrassment we might feel while doing something un-familiar increases when others are present (Kinard, Capella & Kinard 2009). However, this study did not find this effect. No relationship between the choice of a service partner and a non-interactive social presence could be found. This might be due to the fact that for most participants it is the least im-portant feature and therefore is not part of people’s considerations while making the choice. Even if social presence does not influence the type of service partner someone is choosing, it has to be mentioned that in general the absence of others was preferred by participants in this study. That matches previous findings mentioned before. There is also a segment for which it is most important to have no one present around them.
Appendix B: Factor analyses of personality items. (continued) Scree plot of task perception. Scree plot of embarrassment. Appendix C: Results of segmentation analysis for different segment sizes.
