Intelligent Personal Assistants : privacy concerns, perceived risk and other factors influencing intention to use IPA technology

(1)

Master Thesis

Intelligent Personal Assistants: privacy concerns, perceived risk

and other factors influencing intention to use IPA technology.

Student: Guus Aelen

Student number: 11142561

Institution: University of Amsterdam

Track: MSc. Business Administration - Digital Business Email: Guus.aelen@student.uva.nl

Supervisor: dhr. dr. D. (Nick) van der Meulen Submission: 23-06-2017

(2)

2

Statement of originality

This document is written by student Guus Aelen who declares to take full responsibility for the contents of this document.

I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it. The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

(3)

3

This study empirically examines what factors and consumer attitudes influence the acceptance of Intelligent Personal Assistant (IPA) technology in smart homes. Enabled by advancements in the field of Internet of Things and Artificial Intelligence, IPA technology, embodied by devices like Amazon Echo or Google Home, has the potential to revolutionize the way of human-computer interaction. For this study, the UTAUT2 framework was complemented with Concern For Information Privacy (CFIP) and Perceived Risk. Using a multiple regression on a sample of 222 respondents, the model was able to explain 62% of the variance in behavioral intention to use IPA technology. Performance Expectancy,

Hedonic Motivation, Social Influence and Price Value proved to be significant predictors,

substantiating previous findings claiming that Price Value and Hedonic Motivation are important predictors in consumer acceptance of technology. No empirical evidence was found for a significant influence of Facilitating Conditions and Effort Expectancy, which could be due to the relatively young sample. Concern For Information Privacy did not have a direct effect on behavioral intention, but significantly increased a person’s Perceived Risk, which turned out to significantly lower a person’s behavioral intention to use IPA technology (full mediation).

A CFA shows that the CFIP instrument did not perform as expected, with a low factor loading of Concern for Collection compared to the other three constructs of the latent variable. It shows that CFIP should not always be used as a multidimensional construct, and the negative correlation with other UTAUT constructs shows that the influence of Concern

for Collection is more complex and needs further investigation in context of IPA technology.

Key words: Smart assistant, Intelligent Personal assistant, Technology acceptance, virtual assistant, Concern for Information Privacy, Perceived Risk, UTAUT, Zero UI

(6)

6

1 Introduction

The pace by which Internet Technology is changing the way companies work and consumers live their daily life is increasing. Driven by new improved sensoring, wireless technology and other technological innovations, companies equip their products with an internet connection while working to an interconnected network of ‘smart’ related products and services. This emerging development, that goes by the name of the Internet of Things (IoT) has the potential to make supply-chains of companies more efficient and increase convenience and efficiency in the daily life of consumers. According to research of Gartner (2013) the number of ‘connected’ objects will grow to an installed base of 26 billion units by 2020, excluding PCs, tablets and smartphones. This is an exponential growth of 30 times the installed base of 2009, when there were ‘only’ 900 million devices connected to the web. The growth in IoT far exceeds that of any other connected device. To put this in contrast, by 2020 the number of smartphones, tablets and PCs in use will reach about 7.3 billion units. Gartner claims that this rise and the turnover of $300 billion by this time are enabled by the decreased components costs making connectivity a standard feature. Also, the pervasiveness of wireless technology, standardization of protocols, ubiquitous access to internet and consumer familiarity with connected things (e.g. vehicles, smart homes) encourage this (W.-J. Lee & Chong, 2016). IoT, that used to be a fuzzy concept to consumers, is becoming more tangible nowadays. One of the cutting edge applications of IoT technology is the voice-controlled Intelligent Personal Assistant (IPA), such as Amazon’s Echo and Google Home. These devices, enabled by advancements in the field of Artificial intelligence (AI) and Human-Computer interaction, could change the way we interact with technology and take in a central role in our future smart homes, functioning as a voice controlled personal assistant to control smart devices around your house, keep your calendar up-to-date or shop online without the use of a User Interface (UI). Research firm Tractica estimated that 40 million homes will be using digital assistants by 2012 (Bergen & Kharif, 2017). IoT recently received a lot of

(7)

7

attention in practical and academic fields. More than 1400 items were published in 2016, compared to ‘only’ 400 items in 2014 (Web of Science,2017) . The majority of the current research has a technical approach. Research on technologies, privacy and data security are in abundance, while little is known about the consumers that will have to adopt IPA technology.

1.1 Goal and contribution

This study aims to identify what factors and consumer attitudes influence the behavioral intentions to use IPA technology, in order to understand how these factors and attitudes influence future adoption. This study adds to current literature on IoT and technology adoption since it takes a consumer perspective and is (to the best of my knowledge) the first that looks into adoption of IPA technology as one of the most integrating IoT applications for consumers at this time. IPA technology has the potential to change the way humans interact with technology as it enables moving away from touchscreen and other physical User Interfaces. This study is also unique in a way that it takes a consumer-centric approach, in a field where technical research dominates. Practitioners and companies in home industry will benefit by getting a better understanding of what affects consumer adoption, helping them improve the fit between their products and consumer needs. For this study, the following research question is used:

‘’What factors and consumer attitudes influence the acceptance of Intelligent Personal Assistant technology in smart homes?’’

1.2 Research approach

A deductive research approach is used to conduct this research, involving a survey that samples consumers. Based on earlier technology acceptance research, combined with previous studies on Internet of things adoption and similar fields that come close to IPA

(8)

8

functionalities (e.g. E-commerce and smart-home adoption ) and cover focus topics (e.g. privacy), nine hypotheses are suggested. These hypotheses are tested using data from an non-probability sample mainly consisting of students (convenience sample). Data is collected via a web-based survey using Qualtrics analytics.

2 Literature review

For this literature review a combination of business journals and peer-reviewed journals are used. At the start, online business journals like Harvard Business Review (HBR), MIT Sloan (SMR) and California Management Review (CMR) were consulted to identify how IoT and IPA technology relates to today’s business and management practice and identify underlying papers. (trend-)Reports from prominent consulting firms (e.g. McKinsey, PwC, Gartner) provided insight in challenges, developments and terminology that was used afterwards in CataloguePlus (UvA) and Web of Science. No filters for business journals were applied, since a lot of papers are published in the field of computer and information science. Especially peer-reviewed journals and conference proceedings were selected for further reading. Search terms used are included in table 1.

Table 1: Search terms

Initial search terms (HBR,SMR,CMR): Revised search terms (after reading): ‘’Internet of things’’ , ‘’IoT’’, ‘’Personal

assistants’’, ‘’smart home’’, ‘’technology acceptance’’, ‘’Technology acceptance model’’

‘’Internet of things’’ and ‘’adoption’’, ‘’IoT’’ and ‘’adoption’’’, ‘’Intelligent Personal assistants’’, ‘’technology adoption’’, ‘’artificial intelligence’’, ‘’intelligent agent’’, ‘’smart speaker’’

2.1 Internet of Things (IoT)

Most people will say that the internet connects people and enables them to exchange information. More recently, fueled by the use of social media, the internet became a place in which people participate, interact and reside in virtual communities creating their own content. This user-centric view of the internet goes by the name of web 2.0 (Allen, 2009).

(9)

9

IoT, first introduced in 1999 is different from the web 2.0 view since it takes machines and objects with embedded sensors centric view that allow them to communicate autonomously over the internet. Conceptually, the IoT build on three pillars that describe the ability of these ‘smart’ objects: 1) they are identifiable, 2) they communicate, 3) they interact – either among themselves , resulting in a network of interconnected objects (Machine to Machine), or with end-users or other entities in a network (Miorandi, Sicari, De Pellegrini, & Chlamtac, 2012). All these integrated objects create an environment in which consumers are more or less unaware of all the interacting technology surrounding them. This pervasive availability of computers also goes by the name of ubiquitous computing. One of the areas wherein IoT is being used more and more often is within domestic homes. Manufacturers are making devices and appliances around our homes interconnected, giving us control over our domestic environment with the ultimate goal to save cost, live more efficient, enhance functionality or improve quality of live in any other way. This specific market of ‘smart homes’ is expected to grow from $40 million in 2012 to $26 billion in 2019 (Wilson, Hargreaves, & Hauxwell-Baldwin, 2015).

2.2 Risks of IoT

As the spread of connected devices goes on, the paradigm of ubiquitous computing will soon be reality. This includes the domestic environment. Our homes are treasured places. All these smart devices exchange personal data in order to function and integrate into our lives. Data that can, in theory, be used harmfully. This is why security and privacy risks of IoT devices is mentioned as a concern in research (C.-L. Hsu & Lin, 2016; Weber, 2010). Privacy violations like selling data to third parties, identity theft or infrastructure vulnerabilities and the potential use of (domestic) data in digital forensics are topics that lack knowledge and need more regulation (Plachkinova, Vo, & Alluhaidan, 2016). Researchers at this point in time are trying to keep up with the pace of IoT, trying to solve this ‘dark side of IoT’ and its

(10)

10

challenges. At the same time, smart devices are gaining popularity, leaving the untaught or (purposely) unaware consumers with potential large risks.

2.3 Prior research on IoT adoption

We already live in a world with connected vehicles, smart appliances and almost every single electronic device that we buy having an internet connection. Due to the fact that IoT products and services are developing at this moment in time, most research thus far is focused on (technical) challenges like the lack of (open) standards (Miorandi et al., 2012), improving business models (Ives, Blake, 2016; Manyika, 2015) and addressing vulnerabilities and privacy related issues (C.-L. Hsu & Lin, 2016; Plachkinova et al., 2016; Yan, Zhang, & Vasilakos, 2014). It seems like most of this research is pushed by the developers of IoT technology. When it comes to IoT and customer perspective, specifically consumer acceptance and influencing determinants, only a few recent studies were found. A recent study of (C.-L. Hsu & Lin, 2016) provided a framework from the perspective of network externalities and Concern For Information Privacy in order to find out what drives consumers in continued using of IoT services. Based on a literature study they used the predictor variables (PV) Concern For Information Privacy and perceived benefits in combination with the mediator attitude to measure the OV continued intention to use. Their findings show that network externalities (e.g. number of services, perceived compatibility) play a significant role in influencing benefits and thus adoption. Surprisingly, the Concern

For Information Privacy only had a weak (negative) effect on continued intention to use

relative to perceived benefits. Another recent study from Lee & Chong, (2016) also emphasized the importance of understanding consumer acceptance in order to for IoT to become commonly accepted. In this study, the authors used a dual-factor model in order to determine future adoption of smart IoT services and applications. They based their model on Herzberg’s theory (1959), which is ought to be useful for understanding technology services.

(11)

11

They combined this dual-factor theory (avoiding pain and growing psychologically) in the form of satisfiers and dissatisfiers with the technology paradox of Mick and Fournier (1998) to come up with eight PVs influencing satisfaction with IoT (technology attitude) and therefor intention to adopt OV. They found that several satisfiers such as fulfill needs,

perceived enjoyment and technology trust are effective drivers of attitude towards IoT. They

found that their dissatisfiers (e.g. performance ambiguity, perceived incompetence, perceived

chaos) aren’t powerful anti-drivers of IoT service satisfaction. As an explanation, the authors

mention that this could be due to the fact that IoT services are often invisible, making it hard to recognize failure, and adoption is still increasing. They argue that this could result in only few consumers with actual experiences and therefor a weaker ability to connect negative attitudes and dissatisfaction to it. A limitation to this last study is the sample consisting of younger college students, to which an explanation of IoT was presented. It would be interesting to use a more demographically experienced and diverse sample since IoT will affect young and old, and we need to understand motivations of miscellaneous people for IoT services to become ubiquitous.

2.4 Intelligent Personal Assistant (IPA)

Intelligent Personal assistant (IPA) is a fairly new concept for a phenomenon that has been around for ages: Personal Assistants. In literature, IPA are described as ‘’autonomous software agents which assist users in performing specific tasks. They should be able to communicate, cooperate, discuss and guide people’’ (Garrido, Martinez, & Guetl, 2010). The basis of an IPA is formed by a software agent, capable of finding and filtering information, negotiate for services, automate complex tasks, or communicate with other software agents (e.g. devices) to solve complex problems (Czibula, Guran, Czibula, & Cojocar, 2009). Ubiquitous computing and advancements in the field of artificial intelligence, understanding of the semantic web (machine Learning) and speech recognition (capturing the meaning of

(12)

12

spoken language) enabled IPA to become ‘intelligent’. By intelligent is meant that the software agent is able to understand its surrounding (e.g. sensors or communicate with smart objects) and has learning capabilities, resulting in autonomous actions supporting humans. Speech recognition technology enable users to interact with IPA using only their voice, translating natural voice commands into machine-level commands (Santos, Rodrigues, Casal, Saleem, & Denisov, 2016). Another way to understand IPA technology is to look at it as an User Interface. Nowadays we are used to consume information, perform task or control our environment using mobile devices, wearables or laptops. In the beginning, IPA fulfill the same function but without the use of a User interface (UI). This development of not relying on touchscreens and interact with technology in a more natural way also goes by the name of Zero UI. There are different digital assistants currently available. One of the most familiar would be Apple’s Siri, the voice assistant that apple builds into their products. Microsoft named their digital assistant Cortana and Android devices use Google’s assistant. More recently, Amazon introduced their ‘Amazon echo’ speaker and short afterwards Google introduced the ‘Google home’ speaker. These competing devices are currently the most ‘intelligent’ personal assistants, meaning they have the most functionalities, solely using voice commands for interaction. These devices come closest to the definition of IPA and are the most tangible applications of IPA to contemporary consumers.

2.5 Acceptance of Information Technology

Looking into the acceptance of new technology, research draws on different theories and frameworks. A lot of models that seem to be applicable in the case of IoT can be lend from the field of IS. Popular models are the TOE framework (Tornatzky, Fleischer, & Chakrabarti, 1990), UTAUT (Venkatesh, Morris, Davis, & Davis, 2003) and the more popular Technology Acceptance Model (TAM) and Diffusion of Innovation (DOI) (Rogers, 1983; Venkatesh & Davis, 2000). The TOE framework incorporates an organizational perspective. Since this

(13)

13

study focusses on consumers and IoT adoption, this model is not considered. The other three and their applicability are discussed hereafter.

2.5.1 Rogers theory of Diffusion of Innovation (DOI)

The way new innovations (technologies) are accepted and used by different people in society over time is what we call diffusion. According to rogers (1983), whether or not and by which rate innovations are adopted depends on the characteristics of the innovation. The five characteristics of an innovation that determine the rate of adoption by the social system are

(1) Relative advantage: the degree to which an innovation provides value over the idea it supersedes. If an innovation is perceived as advantageous, the higher its adoption rate will be.

(2) Compatibility: the degree to which an innovation fits with existing values,

experiences and needs of potential adopters. It will be more likely for someone to adopt an innovation if this innovation is compatible with situational needs, values and their personal life.

(3) Complexity: the degree to which an innovation is difficult to understand or use. (4) Trialability: the degree to which an innovation can be tested or experimented with. (5) Observability: the degree in which benefits of an innovation are observed by others.

Another influencing factor is how people go through five interlinked stages that he calls the innovation-decision process. These stages consist of (1) how an individual goes from first awareness of an innovation (knowledge), (2) to forming an attitude towards it (persuasion), (3) decides to reject or adopt (decision), (4) to implementation of the innovation (implementation) into (5) reflecting on the decision with preceding expectations (reflection).

(14)

14

2.5.2 Technology Acceptance Model (TAM)

The Technology Acceptance Model (TAM) was first initiated by Davis (1989), and is perceived as one of the most used frameworks in IS to measure technology acceptance (figure 1). It’s based on two

central constructs that determine the attitude and behavior towards a technology, trying to explain the actual intention to use. The two central constructs are perceived ease of use

(EOU) and perceived usefulness (U) later extended by adding two new constructs (TAM2),

being Social Influence and cognitive instrumental processes (Venkatesh & Davis, 2000). Different external factors influencing the central constructs can be incorporated in the model of the researcher, depending on the situation. With more than 700 citations of the initiating article, the TAM model has proven its reliability and perceived usefulness has proven to be a strong (Beta 0.6) driver of Intention to use (Venkatesh & Davis, 2000). However, there are also some studies that showed the TAM model isn’t perfect (Chuttur, 2009). For example, the study of Burton-Jones & Hubona (2006) showed that some external variables have a direct influence on system usage, over and above their indirect effects. As plead by Chuttur (2009), who did an overview of the TAM model, the model reached a saturation point. Future research should combine the strong points of this model, while discarding the weaknesses.

(15)

15

2.5.3 Unified Theory of Acceptance and Use of Technology (UTAUT)

The Unified Theory of Acceptance and Use of Technology (UTAUT) model is the result of a review on user acceptance literature by Venkatesh et al., (2003). The model (figure 2) integrates elements of eight different acceptance models like the Theory of reasoned action, Planned behavior, TAM, DOI and social cognitive theory and others. Venkatesh distilled these theories into the unified UTAUT model, with four core determinants of intention and usage, being (1) Performance Expectancy, (2) Effort Expectancy, (3) Social Influence, (4)

Facilitating Conditions and four moderators of key relationships of these determinants being

(1) gender, (2) age, (3) experience and (4) voluntariness of use. They tested this model and found that it outperforms all preceding models in predicting usage with an adjusted R2 of 0.69 (Venkatesh et al., 2003). The UTAUT model was extended by Venkatesh, Thong, & Xu (2012) into UTAUT2. The authors incorporated three new constructs into the original UTAUT model by adding Hedonic Motivation, Price Value and habit in order to extend and tailor the applicability of the model from employee technology acceptance to a more consumer focused context. Voluntariness of use was removed from the model. In the original UTAUT, utilitarian constructs like Performance Expectancy proved to be a strong predictor of behavioral intention (Venkatesh et al., 2003). Because hedonic or intrinsic motivation proved to be an important perspective in motivation theory (Holbrook & Hirschman, 1982) and is therefore commonly used as important predictor in consumer behaviour research, Venkatesh et al. (2012) decided to incorporate this into the UTAUT2 model. Their results showed that Hedonic Motivation even is a more important driver of Behavioural intention in non-organizational settings compared to Performance Expectancy (Venkatesh et al., 2012).

Price Value is described as the cognitive trade-off between perceived benefits and monetary

costs for using them. Price Value is included since a positive Price Value (benefits outweigh costs) is perceived to have a positive impact on intention thus behavioural intention to use.

(16)

16

experiences’’ (Venkatesh et al., 2012). Finally, the authors expect that the different

Facilitating Conditions of behavioural intention are moderated by age, gender and experience.

(17)

17

2.6 Hypothesis and conceptual model

From the previous literature review it is concluded that the UTAUT2 model seems most applicable for this study, and is therefore used as basis for the conceptual model of this study. The UTAUT2 integrates different proven theories and frameworks on individual user acceptance, and is tailored to fit consumer acceptance by adding Hedonic Motivation Price

Value and Habit. Besides it includes social context in which a technology is adopted while

TAM neglects this. This is perceived important since this research is situated within smart homes, and Social Influence is ought to be important in consumer acceptance. The Habit construct, defined as ‘’the extent to which people perform behavior automatically because of learning’’ (Limayem, Hirt, & Cheung, 2008) is not taken into account in this study, since habit is partly operationalized by prior behavior and learning over time (automaticity), which is perceived as irrelevant when sampling people that most likely do not have any prior experience with IPA technology. Hereafter the different constructs and hypothesis of this study are discussed and operationalized. The conceptual model for this study is visualized in figure 3.

(18)

18 Concern For Information Privacy Effort Expectancy Social Influence Price value Hedonic Motivation Facilitating Conditions Behavioral Intention Performance Expectancy Collection Improper access Unauthorized secondary use Errors Collection Improper access Unauthorized secondary use Errors Perceived Risk H2 + H3

-Figure 3: Conceptual model of this study

2.6.1 Behavioral intention (BI) and actual use behavior

One of the underpinning theories of technology acceptance frameworks is the theory of reasoned action (TRA) by Fishbein and Azjen (Ajzen, 1980). This theory has proven to be useful in explaining and predicting actual behavior of a consumer. It describes that a consumer’s use behavior (actual use) can largely be predicted by analyzing the subjective norms (Social Influence) and attitudes or beliefs which are influencing one’s behavioral intention. The theory was revised by adding perceived behavioral control, meaning the perceived possession of requisite resources to perform a behavior (Ajzen, 1985). In this study, this aspect is incorporated in the construct Facilitating Conditions which is discussed hereafter. Since IPA technology is new to consumers, they are most likely inexperienced with the technology and smart products incorporating it. This means actual use cannot be measured. So in line with other studies on future use of technology (Lee & Chong, 2016; Gao & Bai, 2014; Pavlou, 2003), use behavior is left out and we follow the logic that actual use follows on behavioral intention.

(19)

19

2.6.2 Concern for information privacy (CFIP)

Information privacy can be seen as the right to select what information, when, how and to what extent is known to different people or organizations (Westin, 1967). Nowadays consumers generate a wealth of data that helps companies personalize and tailor their products. Not knowing what happens with personal data and lack of information control gives consumers a vulnerable feeling (Sheng, Nah, & Siau, 2008). This convenience (personalized services) versus privacy (exchange of data) tradeoff goes by the name of the Personalization Privacy Paradox. Literature often treats privacy as a multidimensional (composite) construct. An often used and proven construct is Concern For Information Privacy (CFIP). Smith, Milberg, & Burke (1996) created a valid measurement for CFIP that described how organizations touch upon privacy along four dimensions, being collection, improper access, unauthorized secondary use and errors. A individual is high on CFIP if one perceives that 1) too much personal data is collected (collection), 2) organizations fail to protect access to personal data (improper access), 3) personal data is used for undisclosed purposes (Unauthorized secondary use) or 4) personal data is false or erroneous (Errors) (Stewart & Segars, 2002). Another construct that is used to measure information privacy is the multidimensional construct Internet users’ information privacy concerns (IUIPC) as proposed by (Malhotra, Kim, & Agarwal, 2004). IUIPC is a reaction to CFIP, since

information privacy changed due to the widespread use of internet. This shift in environment plead for specific dimensions tailored to the online environment. Almost resembling CFIP,

IUIPC is conceptualized from three perceptions, being 1) concern of collection of personal

data, which is only perceived as fair if a person is given 2) control over their data, while a person is also 3) informed about the purpose of data collection (Malhotra et al., 2004). This resulted in three (first-order) factors that describe IUIPC: Collection, Control and awareness

(20)

20

stronger with criterion variables and proved to have a better model fit compared to CFIP, researchers still tend to use CFIP to measure Information Privacy Concerns (Belanger & Crossler, 2011). No clear argument for this was given, and because the two instruments are almost interchangeable, the more common CFIP instrument is used in this study.

During the literature review, several studies on the adoption of IoT were found. Hsu & Lin (2016) showed in their study on the adoption of the IOT showed that CFIP has a direct negative effect on Continued intention to use. This is in line with Fishbein’s (1980) Theory of Reasoned Action, since privacy concern is perceived a negative antecedent (Fodor & Brem, 2015; Xu & Teo, 2004) that could influence a user’s attitude towards a technology, and ultimately behavioral intention. More practical, privacy concerns around microphone enabled devices (e.g. digital assistants or smart toys) and the price of convenience are fuel for congresses and skeptic news articles. Other studies in the field of IOT also showed that privacy concerns negatively impacts consumer intention and adoption of technology (C.-L. Hsu & Lin, 2016; C.-W. Hsu & Yeh, 2016; Sheng et al., 2008). Thus, it is expected that CFIP negatively influences behavioral intention.

H1: An individual’s Concern For Information Privacy has a significant negative influence on his/her behavioral intention to use IPA technology.

2.6.3 Concern for information privacy (CFIP) , Perceived Risk (PR) and Behavioral Intention (BI)

The functionalities of IPA technology are built around information of its user. If one wants to utilize all the functions of IPA technology, that person should be willing to give up control and personal data to the IPA service provider and let it interact and control other smart devices (e.g. smart lighting or connected appliances) or integrate with online shopping services like Amazon. Full integration of an IPA device could result in unwanted actions. This is illustrated by hundreds of Amazon’s Echo owners, of which the IPA device ordered a

(21)

21

doll house after a news presenter was talking about ‘how she loved that Alexa orders a doll house for the child’, unintentionally activating and commanding hundreds of Echo devices of

people watching the specific broadcast (Liptak, 2017).

Perceived Risk can be operationalized as ‘’the felt uncertainty regarding possible

negative consequences of using a product or service’’ (Featherman & Pavlou, 2003). The construct has six different facets, like the possibility of malfunctioning (performance risk), monetary loss (financial risk), wasting time (time risk), losing status (social risk), a negative effect on the peace of mind (psychological risk) and the loss of control of personal information (privacy risk). For this study, the last facet is used, which measures overall

Perceived Risk. Risk concerns proved to be an important barrier in acceptance of technology

(Lee, 2009). During the rise of e-commerce at the beginning of this century, both Perceived

Risk and trust (willingness to depend and perception of competence or integrity of others)

proved to be direct influencers on willingness to interact with online merchants (Pavlou, 2003; Van Slyke, Shim, Johnson, & Jiang, 2006). The study of Van Slyke et al., (2006) found that information privacy concerns are salient in this situation. Not directly influencing willingness to transact, but mediated by Perceived Risk. On the other hand, the intangible nature and functioning of IPA technology could also look risky to consumers, implicating that risk perception could play an important role while interacting with IPA technology. Since it is expected that Concern For Information Privacy has a negative effect on

Behavioral Intention, it is expected that this relation is partially mediated by Perceived Risk.

Therefore it is proposed that

H2: An individual’s Concern For Information Privacy has a significant positive influence on his/her Perceived Risk of IPA technology

H3: An individual’s Perceived Risk has a significant negative influence on his/her

(22)

22

2.6.4 Performance expectancy (PE), Effort expectancy (EE) and Behavioral Intention (BI)

The strongest predictor in UTAUT on Behavioral intention is Performance Expectancy (PE). In an IS setting, Performance Expectancy means ‘’the degree to which an individual believes that using the system will help him or her to attain gains in job performance’’ (Venkatesh et al., 2003). The construct is derived from five root constructs, being perceived usefulness (TAM/TAM2), extrinsic motivation (motivational model), job-fit (Model of PC Utilization/MPCU), relative advantage (Innovation Diffusion Theory/IDT) and outcome expectations (Social Cognitive Theory) (Venkatesh et al., 2003) . In this study, the redefined definition from UTAUT2 is used, saying that Performance Expectancy is the ‘’degree to which using a technology will provide benefits to consumer in performing certain activities’’ (Venkatesh et al., 2012). A meta-analytic review of 37 studies using the UTAUT model showed that the impact of Performance Expectancy can be classified as medium with a Zr of 0,53 (Taiwo & Downe, 2013). Another phenomenon that could substantiate the relation between Performance Expectancy and Behavioral Intention in a human-computer interaction situation is that humans suffer from algorithm aversion (Dietvorst, Simmons, & Massey, 2015; Yuksel, Collisson, & Czerwinski, 2017). Humans have a higher intolerance for errors made by algorithms (inseparable to IPA technology) compared to human error. The fact that confidence is lost more quickly if errors are made by technology (no benefits are derived) could plead for a substantial importance of Performance Expectancy in this study. It is expected that

H4: An individual’s Performance Expectancy will have a significant positive influence on his/her behavioral intention to use IPA technology

Another antecedent of Behavioral intention is Effort Expectancy. This is described as the ease of use that is associated with a certain technology. The concept of Effort Expectancy is based

(23)

23

on perceived ease of use (TAM/TAM2) and complexity of a system (MPCU/IDT) (Venkatesh et al., 2003). With both of the constructs being proven predictors within UTAUT(2), it is expected that

H5: An individual’s Effort Expectancy will have a significant positive influence on his/her Behavioral intention to use IPA technology

2.6.5 Social Influence (SI), Facilitating conditions (FC) and Behavioral intention (BI)

Social Influence is derived from subjective norm (TAM/TAM2) , social factors (MPCU) and

image (IDT). All of these labels are based on the notion that an individual’s Behavioral

Intention to use a system is influenced by the judgement or meaning about them of others,

resulting from using a certain technology (Venkatesh et al., 2003). In a consumer setting,

Social Influence is described as ‘’the extent to which consumers perceive that important

others (e.g. family and friends) believe they should use a particular technology’’ (Venkatesh et al., 2003). This would mean that information or encouragement of others influences a user’s Behavioral Intention to use IPA technology. The last decade, it can be concluded that consumers are even more subject to Social Influence because of constant connectivity with peers and their opinion via social media. Consumers share their experience, refer and use review websites to gain knowledge before buying a certain product. Research by Risselada, Verhoef, and Bijmolt (2013) showed that there is a significant positive effect of Social

Influence on adoption of high-technology products, to which IPA technology also belongs.

They also showed that this effect of Social Influence decreases from introduction onwards. As an explanation, Risselada et al. (2013) state that this is due to information substitution; as time goes by more information and common knowledge becomes available about a technology, decreasing the need for opinions or obtaining information from others (Chen, Wang, & Xie, 2011). The fact that users of IPA technology have to interact using their own

(24)

24

voice could also be subjected to Social Influence. At this point in time, the propriety of a human having a conversation with an natural sounding artificial agent is still questionable. Our homes are common places, and people might feel embarrassed drawing attention while their asking their IPA what is in the fridge or to pre-condition the Tesla. Studies on how people use voice-assistants like Apple’s Siri also indicate that the social context and type of information that is transmitted (private versus non-private) influences how people use this technology due to perceived discomfort or the feeling that they are being judged (Moorthy & Vu, 2015). This indicates that the opinion of others influences one’s behavior. In line with these findings, it is expected that a consumer is prone to the technology judgements or opinions of important others, meaning that others have a positive influence on an individual’s behavioral intention to use IPA technology. Therefor it is proposed that

H6: Social Influence on an individual will have a significant positive influence on his/her behavioral intention to use IPA technology

Facilitating Conditions are defined as the extent to which a consumer perceives that

resources exists to support use of a technology. While Facilitating Conditions is directly linked to use behavior (not measured) in the original UTAUT, it is also linked to Behavioral

intention in UTAUT2. This is because facilitating factors (e.g. training) are self-evident in

organizations, but consumers do not always have the same access to facilitating factors and therefore could be more hesitant (Venkatesh et al., 2012). That is why facilitating factors for IPA technology, like online help or available support from a manufacturer, apart from directly influencing use behavior, also influences Behavioral Intention. This construct originally tried to measure to what extent an organizational or technical infrastructure exists to support use of a system. Compatibility, perceived behavioral control (derived from the theory of planned behavior) and technology Facilitating Conditions are therefore seen as root constructs embodied by Facilitating Conditions. This is closely related to a study of Hsu and

(25)

25

Lin (2016), who studied the effect of network externalities on IoT adoption and showed that indirect network externalities like perceived compatibility and perceived complementarity significantly influenced the perceived benefits and thus adoption of IoT services. In a smart home situation, this could implicate that the amount of integrations and (complementary) products or services that an IPA works with are influencing a consumer’s perceived benefits of the technology. It is therefore proposed that

H7: Facilitating Conditions will have a significant positive influence on an individual’s behavioral intention to use IPA technology

2.6.6 Hedonic motivation (HM), Price value (PV) and Behavioral intention (BI)

Hedonic Motivation (HM) and Price Value (PV) are two extensions done by Venkatesh et al.,

(2012) to the original UTAUT to make the model more applicable to a consumer context. Results of this study show that in this context, Hedonic Motivation and Price Value are even stronger predictors of Behavioral Intention compared to Performance Expectancy in the original UTAUT. Intrinsic or Hedonic Motivation, which can be described as the perceived enjoyment or pleasure that is derived from an activity has proven to be an important predictor in Behavioral Intention (Brown & Venkatesh, 2005). Since IPA technology is based on a unique and fun experience of human-computer (two-way) interaction, and voice controlled interaction with a device is fairly new to consumers, Hedonic Motivation will likely play an important role in the tested model. A study of Van Der Heijden (2004) on pleasure oriented information systems supports Hedonic Motivation as important determinant, stating that hedonic systems also focusing on the fun aspect encourage the prolonged usage of a system rather than utilitarian (productivity) use. Another study of Sun & Zhang (2006) proved that

Hedonic Motivation influences Behavioral Intention via perceived ease of use (Effort Expectancy) and found no direct relation. However, since this study had no particular

(26)

26

consumer focus, and this study is interested in the specific hedonic nature of IPA technology in consumer context, it is proposed that.

H8: An individual’s Hedonic Motivation will have a significant positive influence on his/her Behavioral Intention to use IPA technology

In organizational settings, workers do not bear the monetary costs of an information system. However, this is the case for consumers adopting new technology. It is evident that the price or cost of a technology will influence it’s use. Price Value is therefore defined by Venkatesh et al., (2012) as a ‘’consumers’ tradeoff between the perceived benefits of the application and the monetary cost for using them’’ (Dodds, Monroe, & Grewal, 1991). As soon as benefits of a technology outweigh the monetary costs, Price Value of a technology is perceived as positive and positively impacts Behavioral Intention (Venkatesh et al., 2012). Therefor it is proposed that

H9: Price Value will have a significant positive influence on an individual’s Behavioral

(27)

27

3

Data collection

This study uses quantitative, primary survey data in order to understand how the different constructs affect the behavioral intention of consumer to use IPA technology. In this chapter the reliability of the items and data collection method is explained.

3.1 Instrument development

In the previous paragraph the constructs were operationalized. Before spreading a survey, it is important that a construct is measuring what is claims to measure. This is also called the content validity of the constructs one is interested in (Saunders, Lewis, & Thornhill, 2012). To ensure this, it is important to use validated items that are altered to a minimum while adapting them to the context of IPA technology. The adjusted and original items can be found in appendix 1.Since all items are stemming from English studies, and some respondents had Dutch as their first language, the questionnaire was translated to Dutch for convenience. To prevent translation errors, questions that were difficult to translate were put into Google Translate, that in a few cases suggested a better translation. Some of the translated questions were translated back using the same service to check whether the translation was done accurate. As described in table 2, a multi-item method is used with 3-4 questions per construct. The reliability column shows that Composite Reliability (CR) exceeds the threshold of .70 and the Average Variance Extracted (AVE) exceeds the threshold of .50 (Hair, 2010), showing the measures proposed for this study are internally consistent and reliable enough for use. Except the CR collection does not exceed this threshold, which is close to .7 (.68). While there are studies claiming that CR above .60 is also acceptable (Tseng, Dörnyei, & Schmitt, 2006), this slightly lower reliability score will be kept in mind during further analysis.

In line with the popular technology acceptance models, a seven point Likert-scale ranging from 1 (strongly disagree) to 7 (strongly agree) is used for most of the items. Perceived Risk

(28)

28

was measured using a different scale as suggested by Sitkin & Weingart (1995). See item in appendix 1for more information.

Table 2 : Overview of previous construct items and reliability

3.2 Sampling strategy

The primary data for this study was gathered using Qualtrics. This is the recommended online survey software licensed by the University of Amsterdam. The entity that is studied, and the population that it should describe, are normal consumers, not belonging to a certain (age)group, geographical unit or characterized by a certain social aspect. This makes the unit of analysis an individual consumer. The final survey consisted of 46 questions, including demographic data and optional comments. To reduce social desirable answers, the respondents were prompted a message upfront saying that there are no right or wrong answers and data is treated anonymously. To enhance the completion rate, a boat trip and Construct Items Original reliability Retrieved from

Concern for Information Privacy a

- Unauthorized secondary use - Improper access - Errors - Collection 4 CR=.94 CR=.95 CR=.94 CR=.68

(C.-L. Hsu & Lin, 2016; Smith et al., 1996; Stewart & Segars, 2002)

Behavioral intention (BI) 3 ICR (AVE) = .94 (.82) (Davis, 1989; Venkatesh et al., 2003) TAM/UTAUT

Performance Expectancy (PE) 3 ICR (AVE) = .88 (.75) (Venkatesh et al., 2003) TAM

Effort Expectancy (EE) 4 ICR (AVE) = .91 (.74) (Davis, 1989; Venkatesh et al., 2003) / TAM UTAUT

Social Influence (SI) 3 ICR (AVE) = .82 (.71) (Venkatesh et al., 2012) / UTAUT2 Price value (PV) 3 ICR (AVE) = .85 (.73) (Venkatesh et al., 2012) / UTAUT2 Hedonic motivation (HM) 3 ICR (AVE) = .86 (.74) (Venkatesh et al., 2012) / UTAUT2 Facilitating conditions (FC) 4 ICR (AVE) = .75 (.73) (Davis, 1989) / TAM

Perceived Risk (PR) 4 (Featherman & Pavlou, 2003; Sitkin &

(29)

29

glider flight were raffled amongst complete entries. To prevent fatigue and dishonesty, answering a question was not mandatory and respondents were able to drop out at any time. Since respondents were likely unexperienced with the technology, a picture of two Intelligent Personal Assistants was showed, together with a brief introduction on the technology’s functionalities and its applications. After the survey was pre-tested among a group of trusted respondents, the somewhat difficult Likert rating of one of the items of Perceived Risk (insignificant risk – significant risk) was improved (low risk-high risk). Pre-test entries were removed from the final dataset. The survey was spread via Facebook, LinkedIn, family, friends and through a direct mail to around 400 students of a quantitative data workshop, making the dataset a convenience sample. Since the amount of younger consumers (students) present is likely to be high, caution should be taken in generalizing the results. Over-representation of certain groups is typical drawback of convenience sampling (Saunders et al., 2012). The data was gathered between the 9th and 24th of May 2017. The survey that was used in this research can be found in appendix 2.

(30)

30

4 Data analysis and findings

A total of 258 people entered the survey, which took on average 7.5 minutes to complete. People that abandoned the survey, did this in the beginning after answering only one or two questions. Because of this, it was decided to delete all incomplete responses together with empty responses (n=30). Three responses were deleted because the response seemed dishonest (e.g. same answer 18 times in a row) in combination with a unlikely response time under two minutes. No counter-indicative items were included. After further screening, a total of 222 cases were valid for further analysis. SPSS 22 and SPSS AMOS were used to analyze the collected data.

4.1 Descriptive measures

As described In table 3,the majority of the respondents had the Dutch nationality (94.1%). Of all respondents, there were 109 female and 103 male respondents. These accounted for respectively 49.1% and 46.4% percent of the total, implicating a good gender balance. The age of the respondents lay between 16-63, with most of the respondents in the age category of 20-29 (77.9%).

Table 3: Descriptive sample measures

Variable (N=222) % Nationality Dutch Non-Dutch Not provided 209 10 3 94.1 4.5 1.4 Gender Female Male Not provided 109 103 10 49.1 46.4 4.5 Age category 20 or under 20-29 30-39 40-49 50 or over Not provided 4 173 16 2 25 2 1.8 77.9 7.2 0.9 11.3 0.9

(31)

31

4.2 Reliability and validity

After skipped questions were recoded into -99 to exclude these from the analysis, several assumption checks were performed to check for normality and outliers in the data before performing any parametric tests. Assumptions specific to multiple regression are discussed hereafter. It was concluded that skewness for all variables was well within the acceptable range of -2 , 2 (Field, 2013). The most extreme value was found for the latent variable

Concern for Secondary Use (-1.56) which is part of the second order construct Concern For Information Privacy. The skewness for this (summated) second order construct was -0.52.

Looking at Q-Q plots, most of the data laid close to the indicator line of a normal distribution. Overall, around 5 data points could be considered as ‘outlier’. It was decided to not remove these points because of the nature of the Likert Scale. The answers in other questions seemed honest, and it did not seem right to delete data because a person ended ‘lower’ or ‘higher’ on a scale with a small 1-7 spectrum. According to the central limit theorem (Field, 2013), which claims that scale means approach a normal distribution as sample size increases, the decent sample size (N=222) also indicates that the scale means are approximately normally distributed. While this is not hard proof, it does substantiates the other indicators which look in order.

Opposite to the original UTAUT2 constructs, of which the factor structure is grounded in different theories, the Concern For Information Privacy instrument is a construct that has not been used in acceptance studies that often. Former studies showed that the instrument can be treated as a second order construct (C.-L. Hsu & Lin, 2016; Stewart & Segars, 2002).However, Stewart & Segars (2002) also mention that the underlying constructs of Concern For Information Privacy and their applicability should be further investigated in light of emerging technologies and different research contexts. Ongoing trends like the Internet of Things and the increased awareness of consumers on how businesses benefit from- and use their personal data makes it difficult to fully understand what factors influence

(32)

32

consumer attitudes towards information privacy. Since this study uses a multiple regression to answer the hypotheses, and the model is not tested as a whole as one would do with Structural Equation Modelling (SEM), a separate Confirmatory Factor Analysis (CFA) was executed upfront using SPSS AMOS to verify the structure and test whether Concern for

Secondary Use, Improper Access, Errors and Collection actually are the underlying factors in

context of this study as theory would suggest. Results are shown in table 4.

Table 4 : Results from CFA with factor loadings, composite reliability and average variance extracted

In order to decide whether we can use the outcome of the CFA model, the so called Goodness Of Fit (GOF) is tested. The RSMEA of .08 implicates ‘reasonable fit’ (Bentler & Bonett, 1980). Preferably, this number should be <.05 to be considered as a perfect fit (Hair, 2010). The CFI fit indice that preferably exceeds .95 for good models, was smaller (.90). All of the Composite Reliability (CR) scores exceeded the .70 threshold, indicating the summated scales are reliable (Fornell & Larcker, 1981). The Average Variance Extracted (AVE) exceeds the threshold of .50 for almost all variables, indicating sufficient convergent validity. The common CFIP factor only seemed to explain .40 of Concern for Secondary Use, what could be due to a lower loading of two of the four items within this factor (see figure 4).The low factor loading of Concern for Collection (.16) is also something that stands out from the analysis. It indicates that this factor has a weak effect on the CFIP construct, and that it possesses less similarities with the other three factors explaining CFIP. Based on the CR of

Relation Factor loading CR AVE

CCOL CFIP .16 .83 .55

CSU CFIP .75 .70 .40

CERR CFIP .57 .84 .57

(33)

33

CSU and CCOL (.70, .83) and the fact that the model fit indices do not directly imply for a perfect model fit, the structure of the CFIP construct is kept as in theory. This decision is also endorsed by an extra Principal Component Analysis that was done, indicating that extracting one factor was the best solution. The Eigen Value (EV) dropped below the threshold of 1 when two or more factors were extracted. While the total explained variance increased from 47% to 71% with two factors, a visual inspection of the Scree plot clearly pleaded for one

factor. Substantial support can also be found in the correlation matrix (table 5) discussed hereafter. In line with the theoretical foundation, the underlying factors of the CFIP scale were summated into one construct for further analysis, with a reliable Crohnbach alpha

).

(34)

34

4.3 Correlations

Table 5 shows the means, standard deviations (SD), reliability and bivariate correlation for the independent, dependent and control variables. Keeping the confounding variable problem and direction of causality in mind, a quick glance at the matrix shows that there are interesting correlations going on. Statistically significant correlations (p<.05) with r= .4 (moderate) or correlations relevant for the focus of this study are discussed.

Looking at the latent Concern For Information Privacy instrument (13), the results show strong correlation (r= [.64, .69] , p=<.01) with the four underlying variables of the It also shows a negative correlation with Performance Expectancy (r= -.19, p=<.01) , Social Influence (r= -.24, p=<.01), Hedonic

Motivation (r= -.24, p=<.01), Price Value (r= -.19, P=<.01) and Behavioral Intention (r= -.22;

p=<.01), what could be a first hint that CFIP plays a role in the adoption of IPA technology.

CFIP is positively correlated with Perceived Risk (r= .32, p=<.01), indicating that one’s Perceived Risk increases when a their Concern For Information Privacy becomes stronger.

The weak negative correlation of CFIP on Behavioral Intention (r= -.22, p=<.01) and moderate to strong correlation between Perceived Risk and Behavioral Intention (r= -.46 , p=<.01) form a hint that a significant mediating effect could be present over a direct effect. This effect will be further discussed in the next chapter. Perceived Risk also shows a

moderate negative correlation with Performance Expectancy (r= -.37, p=<.01), Price Value (r= -.39; p=<.01) and strong negative correlation with Hedonic Motivation (r= -.47, p=<.01). Furthermore, it can be concluded that Hedonic Motivation, a factor claimed to be relevant especially for consumer acceptance of technology (Venkatesh et al., 2012), is indeed strongly correlated with Behavioral Intention to adopt IPA technology (r=.66, p=<.01). While this study also claimed that Hedonic Motivation is an even stronger predictor of Behavioral

Intention over Performance Expectancy, both constructs have the same strong positive

(35)

35

fun aspect and benefits of IPA technology play an important role in the adoption of the technology. The study of Venkatesh et al. (2012) also suggest that if the benefits of IPA technology outweigh the monetary costs, Price Value of the technology is seen as a positive factor and therefor positively influences Behavioral Intention. Looking at the correlation matrix, these findings are reflected in the strong positive correlation between Price Value and

Behavioral Intention (r=.55, p=<.01). But also the moderate positive relation with Performance Expectancy (which include perceived benefits) (r=.43, p=<.01) and Hedonic Motivation (r=.54, p=<.01) indicates that Price Value plays an important role in intention to

(36)

36

Table 5: Correlation matrix

Correlation matrix of all variables

Mean (SD) 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 1. CCOL a 5.01 (1.21) (.80) 2. CSUa_{6.26 (0.78)} _.17* _(.65) 3. CIAa _{6.32 (0.68) .18}** _.54** _(.83) 4. CERa_{5.31 (0.98)} _.07 _.32** _.34** _(.82) 5. PE 4.55 (1.30) -.35** _-.04 _.01 _.00 _(.88) 6. EE 5.44 (0.95) -.23** _.17* _.13 _-.03 _.22** _(.88) 7. SI 3.08 (1.37) -.24** _-.12 _-.21** _-.05 _.38** _-.02 _(.92) 8. FC 4.99 (0.99) -.20** _.10 _.13 _-.07 _.19** _.53** _.09 _(.71) 9. HM 4.84 (1.33) -.42** _.03 _-.04 _-.06 _.61** _.31** _.36** _.27** _(.87) 10. PV 5.00 (0.97) -.30** _-.05 _.00 _-.05 _.43** _.15* _.34** _.22** _.54** _(.81) 11. BI 3.58 (1.51) -.35** _-.07 _-.05 _-.02 _.66** _.20** _.50** _.27** _.66** _.55** _(.92) 12. PR 3.83 (0.92) .55** _.05 _.06 _.02 _-.37**_-.24**_-.27**_-.27**_-.47**_-.39**_-.46**_(.71) 13. CFIP 5.72 (0.60) .64** _.69** _.69** _.64**_-.19** _-.04 _-.24** _-.06 _-.24**_-.19**_-.22**_.32** _(.79) 14. Genb _{0.51 (0.03)} _.12 _.24** _.18* _.23** _-.03 _-.14* _-.05 _-.19** _-.10 _.05 _-.13 _.09 _.28** Note: Item reliability (Alpha

Improper Access; CER: Concern for Errors; PE: Performance Expectancy; EE: Effort Expectancy; SI: Social Influence; FC: Facilitating Conditions; HM: Hedonic Motivation; PV: Price Value; BI: Behavioral Intention; PR: Perceived Risk; CFIP: Concern For Information Privacy

a _{is part of latent variable} b_{Gender dummy female}

*. Correlation is significant at the 0.05 level (2-tailed). **. Correlation is significant at the 0.01 level (2-tailed).

(37)

37

4.4 Analysis

In the next section, the valid data and statistical procedures are further presented. The role of

Perceived Risk is further investigated using SPSS PROCESS. Afterwards, a multiple

regression is used in SPSS to better understand the relationship between the independent factors and Behavioral Intention. An alpha level of .05 is used for all statistical tests. Before trying to explain how the independent variables influence Behavioral Intention and results are compared with the previously drafted hypotheses, it is important to look at the assumptions of a multiple regression. As mentioned in the data collection section, all of the independent and dependent variables are measured on a quantitative scale (scale means), making them suitable for regression analysis. The correlations at the beginning of the analysis section already showed that there were now indicators (r=>.8) (Field, 2013) for perfect multicollinearity. This is highly unwanted, since it would scramble the results of individual predictors and their effect on the outcome variable. In addition to this r=>.8 ‘ball park’ method, the Variance Inflation Factor (VIF), which is a collinearity diagnostic that should be below a value of 10 (Myers, 2000), was calculated for all predictors. With the VIF ranging between 1.17 and 2.20 and tolerance factors all >.46, no reason to concern for multicollinearity was found. Another factor that could blur or give false statements about significance of relations is autocorrelation. This is the case when the residuals terms of different observations over time correlate with each other. To exclude autocorrelation and test for this assumption, the Durbin-Watson test was done. This value should preferably be between 0 and 4, with a value of 2 meaning that the residuals do not correlate with each other (Field, 2013). The Durbin-Watson test gave a value of 1.94, showing that also this assumption was met.

Since a multiple regression is ran, it is of the utmost importance to build a model based on theory. While one is able to add all of the predictor variables in a SEM study at once, predictors should be carefully chosen when conducting a multiple regression.

(38)

38

‘Overfitting’ a regression model, a situation in which a large number of random predicting variables are added in the hope that they are relevant, is an often heard statement (Field, 2013). On this point, a strong theoretical framework and usage of constructs that have theoretically proven to be of importance is key. Naturally, during the literature review, only relevant constructs in light of IPA technology and technology acceptance were taken into account and should all be tested in the regression.

Since this study has specific interest in the role of Concern For Information Privacy on the Behavioral Intention to adopt IPA technology, and it is expected that this relation is partially mediated by Perceived Risk, the multiple regression was preceded by a test for mediation using SPSS PROCESS. The results are shown in table 6 and figure 5.

Table 6: Test for mediation using Process using ‘model 4’, showing direct and indirect effect sizes (N=221)

Consequent

PR (M) BI (Y)

Antecedent Coeff. SE P Coeff. SE P

CFIP (X) A1 .487 .097 < .001 C1 -.202 .158 .201

PR (M) --- --- --- B1 -.716 .103 <.001

constant I1 1.041 .563 0.065 I2 7.475 .866 <.001

R2_{= .102} _R2_{= .221}

F = 24.77 , P<.001 F = 30.90 , P<.001

Effect SE P LLCI ULCI

Direct effect C1’ -.202 .158 >0.05 (.201) -.513 .109 Total effect C1 -.551 .165 < .001 -.875 -.226

Boot SE Boot

LLCI Boot ULCI

(39)

39 CFIP BI PR .487*** -.716 *** A1 B1 C1

Figure 5: Effect of CFIP with the direct and indirect unstandardized coefficients (b) via mediator Perceived Ris. *. Coefficient is significant at the 0.05 level (2-tailed).

**. Coefficient is significant at the 0.01 level (2-tailed). ***. Coefficient is significant at the 0.001 level (2-tailed)

As illustrated in table 6, the effect of CFIP on PR A1 shows that for every 1 unit increase of CFIP, a person also positively increases in Perceived Risk by .487 units (unstandardized

coefficient). The positive relation indicates that if a person is estimated to be higher in CFIP, that person’s Perceived Risk also increases. This effect is statistically different from zero, with p = <.001 and t=4.977, within the 95% confidence interval of .294 to .680. This means that H2 is supported.

The negative effect B1 = -.716 indicates that two people who experience the same

level of CFIP, but differ by one unit in Perceived Risk, are estimated to differ -.716 in

Behavioral Intention. B1 is negative, what indicates that a person higher in Perceived Risk is

also lower in Behavioral Intention. This effect is statistically different from zero, with a t= -6.94, p= <.001 and a 95% confidence interval between -.919 and -.513. This means that H3 is supported.

The negative indirect effect a1b1 = -.349 indicates that two persons that relatively

differs one unit in CFIP, defer approximately by -.349 in Behavioral Intention, because these people are higher in Perceived Risk, what results in a decreased Behavioral Intention to use IPA technology.

Intelligent Personal Assistants : privacy concerns, perceived risk and other factors influencing intention to use IPA technology

Master Thesis