The price of privacy:
What price to ask and who to target for a cloud storage
service that offers full privacy protection.
J.H.H. Kral
11-01-2016
2
Inhoudsopgave
1. Introduction ... 3
2. Literature review and theoretical framework ... 6
2.1 Cloud storage services ... 6
2.2 Privacy ... 7 2.3 Willingness-To-Pay ... 9 2.4 Privacy concerns ... 10 2.5 Socio-demographics ... 11 2.6 Mediated moderation ... 13 2.7 Interaction effects ... 15
2.8 Characteristics of cloud storage services ... 15
2.9 Conceptual model ... 17 3. Methodology ... 18 3.1 Survey design ... 18 3.2 Sample size ... 19 3.3 Statistical methods ... 19 4. Results ... 20 4.1 Descriptive statistics ... 20
4.2 Quality of scale for privacy concerns... 21
4.3 Choice-Based Conjoint analysis ... 25
4.4 Testing the hypotheses ... 31
5. Conclusion ... 37 6. Discussion ... 39 6.1 Managerial implications ... 39 6.2 Limitations ... 40 6.3 Future research ... 40 References ... 42 Appendices ... 46
Appendix 1: Factoring questions about privacy concern: Exploratory correlation matrix ... 46
Appendix 2: Factoring questions about privacy concern: Measures of one-factor solution ... 47
3
1. Introduction
In the last decades the world has moved to a digitized society faster than anyone could have predicted. This comes with great advantages and opportunities. Nowadays, the world is tied together by the Internet, making global communication and integration easier than ever. Think about cloud storage, as a part of cloud computing, which potentially makes local storage of data unnecessary and makes it possible to access your data anywhere you want. But although governments, companies, and consumers all benefit from this, there are some serious threats to watch out for. The digitization of society and the rapid increase in Internet use (read: Internet dependence) resulted in the fact that personal information is much easier to come by for both governments and companies. Add to that the increasing hunger for data most large companies nowadays have and it is easy to imagine that the privacy of information becomes harder and harder to retain for consumers.
A remarkable quote of which no one knows who said it first, but that is widely known now, is that; “If you’re not paying for the product, you are the product”. The business models of the biggest cloud storage services, e.g., Dropbox and Google Drive, rely on this principle. These companies let you use their services for free, but thereby giving themselves rather insane rights as of what they may do with your stored data and personal information. And these rights go a lot further than most people think. Namely, it is common practice for free consumer cloud storage services not to offer any service guarantees, to assume no liability for any data loss, and to reserve the right to disable accounts without reason or prior notification, as well as to change or stop providing the service at any time (Ion et al., 2011). The Google Terms of service also state that;
“Google reserves the right (but shall have no obligation) to pre-screen, review, flag, filter, modify, refuse or remove any or all Content of any Service.”
4 According to research done by Ion et al. (2011) there is a market for such a service. They found that 79 percent of respondents would pay 20 USD per year to a company whose policy says they will not sell any personal information. With a probability bordering on certainty it can be said that a cloud storage service like this is most suitable for consumers that have at least a moderate level of privacy concerns. To get a clearer picture on which consumers this service should be targeted on, it can be interesting to see which consumer characteristics have a substantial impact on the level of privacy concerns.
In privacy surveys since 1978 it appeared that age was a more important factor in explaining attitudes towards privacy than other socio-demographic characteristics, such as gender, income, region, education or race (Westin, 1991). Westin found a positive effect of age on the level of privacy concerns. Sheehan (2002) and Zukowski and Brown (2007) came to the same conclusion. Therefore, age seems to be a good starting point for determining which consumers to target for this new cloud storage service. Although by some researchers age is believed to be the most influential socio-demographic characteristic on privacy concerns, the picture will be far from complete when other socio-demographics are being neglected. Therefore, the focus of this paper will not only be on age but also on gender, household income, and educational background.
It is believed that women are more concerned about their privacy than their male counterparts (Dutton & Meadow, 1985; Youn et al., 2008). The relationship between household income and privacy concerns is somewhat under researched. In their unpublished report for the Office of Technology Assessment in the US Congress, Dutton and Meadow (1985) found that people with a higher income are less concerned about their privacy. Although this research is quite old and focused more on the public perspective on government information technology, it seems reasonable to take into account household income when trying to pinpoint which consumers to target. Especially since the consumer has to pay for this new cloud storage service.
5 In the recent years there came increasingly more attention for information privacy, cloud computing, and a combination of both. However, to the best of my knowledge nobody has yet tried putting a price on a cloud storage service offering full privacy protection, let alone examining which consumers to target. With the following paper I would like to fill that gap.
This is the field this paper will be focusing on. Firstly, I would like to determine what effect the level of privacy protection has on the adoption likelihood of such a service. Followed by determining the price consumers are willing to pay for it. Furthermore, I hope to paint a picture of the influence certain socio-demographic characteristics have on the level of privacy concerns in order to determine on which consumers to target the new service. Finally, I will try to explain the mediating role of the level of privacy concerns of a consumer in the relationship between socio-demographic characteristics and the main effect.
Although there are already some other cloud storage service solutions (e.g. SugarSync) that offer an improved privacy policy (mostly) against a monthly fee, they still deny any liability for loss of personal data. Also, these services seem rather expensive for what they are offering, especially compared to the 20 USD people were willing to pay on a yearly basis in the experiment done by Ion et al. (2011). With regards to the outcomes of this paper I hope I can say something about the feasibility of introducing this service, as well as creating a better understanding as to which consumers to target and what price to ask them.
6
2. Literature review and theoretical framework
In this part I will try to give an overview of the literature regarding the concepts and relationships I propose in this paper. Some key concepts will be given a definition and the proposed relationships will be substantiated. Concluding, I will present the hypotheses that will be tested along with the corresponding conceptual model.
2.1 Cloud storage services
Cloud storage services are part of the overarching concept of cloud computing. The National Institute of Standards and Technology (NIST) defines cloud computing as follows (Mell & Grance, 2011):
“Cloud computing is a model for enabling ubiquitous, convenient, on-demand network access to a shared pool of configurable computing resources (e.g., networks, servers, storage, applications, and services) that can be rapidly provisioned and released with minimal management effort or service provider interaction. This cloud model is composed of five essential characteristics, three service models, and four deployment models.”
According to the NIST these essential characteristics are (Mell & Grance, 2011);
On-demand self-service. A consumer can unilaterally provision computing capabilities, such as server time and network storage, as needed automatically without requiring human interaction with each service provider.
Broad network access. Capabilities are available over the network and accessed through standard mechanisms that promote use by heterogeneous thin or thick client platforms (e.g., mobile phones, tablets, laptops, and workstations).
Resource pooling. The provider’s computing resources are pooled to serve multiple consumers using a multitenant model, with different physical and virtual resources dynamically assigned and reassigned according to consumer demand. There is a sense of location independence in that the customer generally has no control or knowledge over the exact location of the provided resources but may be able to specify location at a higher level of abstraction (e.g., country, state, or datacenter). Examples of resources include storage, processing, memory, and network bandwidth.
Rapid elasticity. Capabilities can be elastically provisioned and released, in some cases automatically, to scale rapidly outward and inward commensurate with demand. To the consumer, the capabilities available for provisioning often appear to be unlimited and can be appropriated in any quantity at any time.
7 monitored, controlled, and reported, providing transparency for both the provider and consumer of the utilized service.
The service models the NIST distinguishes are: Software as a Service (SaaS), Platform as a Service (PaaS), and Infrastructure as a Service (IaaS). The different deployment models they name are: Private cloud, community cloud, public cloud, and hybrid cloud.
According to Horrigan (2008), by 2008, 69 percent of all Internet users had either stored data online or had used a web-based software application. In addition, the market for cloud computing is likely to expand even more in the nearby future. In the beginning of the current decade, Anderson and Rainie (2010) studied the future of cloud computing for the Pew Research Center. They found that a solid majority of technology experts and stakeholders participating in the fourth Future of the Internet survey expect that by 2020 most people will access software applications online and share and access information through the use of remote server networks, rather than depending primarily on tools and information housed on their individual, personal computers (Anderson & Rainie, 2010). Furthermore, these experts said that, in this decade, cloud computing will become more dominant than desktop computing. According to Anderson and Rainie (2010), the biggest driver of the growth of cloud computing is cloud storage. Data is moving from user-owned desktops and laptops to dedicated online storage systems, e.g. Dropbox and Google Docs (Ion et al., 2011).
The use of cloud storage services in itself brings about some great benefits, such as continuous availability of data anywhere, anytime, and easy sharing of pictures and documents with friends and family. Also, if used for purposes of backing up data, it relieves the burden of self-managing replication and data backups (Ion et al., 2011).
However, without appropriate security and privacy solutions designed for clouds, this potentially revolutionizing computer paradigm could become a huge failure (Takabi et al., 2010). According to several surveys among potential cloud adopters, security and privacy are the primary concerns hindering its adoption (Bruening & Treacy, 2009).
2.2 Privacy
8 They found that over a time span of 20-plus years the percentage of people that were “very concerned about threats to privacy” increased from 31% to 61% (1978 – 2000) (Regan et al., 2013).
Users get bombarded with privacy policies and terms of service, from all kinds of services and products, they have to agree to. This is probably the main reason that users don’t read them and believe they have more rights and guarantees than what these documents actually grant them (Ion et al., 2011). Still, this is quite striking when taking into account the number of people that apparently are very concerned about their privacy. It seems that most people are concerned about their privacy but feel it is too much work to control for it when agreeing to all kinds of privacy policies. It is interesting to see how ignorant most people are when it comes to something they say they care about. A related concept is the so called privacy paradox, which will be explained shortly.
In a sense, the ignorance of most people is not that surprising. In their study on an economic (re-)evaluation of privacy, Van Aaken et al. (2014) state that, “privacy is a fuzzy concept”, even for theorists in this field. Regan et al. (2013) support this by stating that privacy is an enormously complicated and broad concept, and it is nearly impossible to phrase a question to capture its nuances. People often talk about privacy in terms of ‘infringements’, ‘violations’, or ‘intrusions’. It seems that privacy is conceived as something one loses rather than possesses, which is probably why people tend to become aware of privacy when they lose or lack it. However, one can give up privacy voluntarily and may even do so in exchange for something else. In many instances, people trade privacy for money (Van Aaken, 2014). It can be argued that this is also the case when consumers adopt free cloud storage services. They trade their privacy for a free service. Van Aaken et al. (2014) conclude their article by stating that privacy is not a taste or preference. It is a specific form of freedom and has intrinsic rather than instrumental value only.
Nevertheless, there are a lot of researchers that have tried conceptualizing privacy, either complementing or contradicting each other. Clarke (1999) identified four distinct dimensions of privacy: privacy of a person, behavior privacy, communication privacy, and data privacy. Bélanger and Crossler (2011) tried to update this view by arguing that, given the digitization of information and communication, personal communication and data privacy can be merged into the information privacy construct. Thereby recognizing that the concept of privacy has to be updated due to the increasing digitization we experience nowadays. A rather old, but still applicable definition of information privacy is that information privacy refers to the claim of individuals, groups, or institutions to determine for themselves when, how, and to what extent information about them is communicated to others (Westin, 1967).
9 (Pavlou, 2011; Norberg et al., 2007). Bennett (1995) explained the privacy paradox by stating that privacy is not absolute, and it can be assigned an economic value based on economic principles, such as a cost-benefit calculation. Related to this is the concept of privacy calculus. Hereby, consumers are willing to share personal information to obtain certain benefits (Ackerman, 2004). It is interesting to see if this concept also works the other way around. That is, are people willing to pay for retaining their privacy?
Privacy requirements for consumer cloud storages differ from those of companies. Users are less concerned about some issues, such as guaranteed selection of data, country of storage and storage outsourcing, but are uncertain about using cloud storage (Ion et al., 2011). The results of Ion et al. (2011) further suggest that end-users consider the Internet as intrinsically insecure and prefer local storage for sensitive data over cloud storage. However, users desire better security and are ready to pay for services that provide strong privacy guarantees (Ion et al., 2011).
2.3 Willingness-To-Pay
Willingness-to-pay is defined as the maximum price a buyer accepts to pay for a given number of goods or services (Kohli & Mahajan, 1991). However, in the case of calculating the WTP for privacy protection, it might be more useful to build on work done by Grossklags and Acquisti (2007). They came up with two variants of WTP: willingness-to-protect and willingness-to-sell. Willingness-to-sell is the willingness to accept a proposal to sell information. Willingness-to-protect is the willingness to pay for protecting information (Grossklags & Acquisti, 2007). The latter, in particular, seems to be very useful and applicable with regard to the subject of this paper. Therefore, in the rest of this paper, when talking about WTP I am actually talking about the concept of Grossklags and Acquisti (2007), named willingness-to-protect. Interesting to note is that in all scenarios they tested, they found the average willingness-to-sell to be dramatically higher than the average willingness-to-protect. This might illustrate part of the privacy paradox. When selling their privacy, people appear to value their private information very much, but when there are costs related to protecting their privacy they value it a lot less.
Building on previous paragraphs it can be assumed that people are willing to pay money for a cloud storage service that provides strong privacy guarantees. Building on that, it can be hypothesized that the WTP and adoption rate get higher when the level of privacy protection of the cloud storage service gets higher. Therefore, my first hypothesis is:
10
2.4 Privacy concerns
In the next few paragraphs I will hypothesize how different socio-demographic characteristics affect the relationship between the level of privacy protection and the WTP and adoption rate of a cloud storage service. For creating a clear understanding I will first talk about the hypothesized underlying concept of their effect, privacy concerns.
Many researchers have adopted the definition of privacy concerns as Smith et al. (1996) have defined it. They have defined privacy concerns as focusing on the concerns individuals have with the information privacy practices of organizations (Smith et al., 1996). Others have applied the definition of Malhotra et al. (2004). They have defined privacy concerns more broadly as the individual’s subjective views of fairness within the context of information privacy (Malhotra et al., 2004). In this paper the definition of Smith et al. (1996) is used.
As mentioned earlier, consumers are getting increasingly concerned about their privacy (Regan et al., 2013). Big part of this growing concern is the rapid digitization in modern times. Especially the Internet raises concerns about the privacy of information (Bowie & Jamal, 2006). The Internet offers immense possibilities to collect and trade private information. Organizations are enabled to collect and combine personal information with the help of several tools, like web analytics software and trackers. This helps customize the Internet and offers conveniences and privileges to users, such as access to particular contents or personalized services. In return users grant them personal information and give up part of their privacy. A lot of the time users are unaware of this trade (Van Aaken et al., 2014). In recent years privacy concerns have been extensively researched. Findings suggest that privacy concerns influence individuals’ acceptance of technology, such as their intentions to purchase online (Malhotra et al., 2004; Smith et al., 1996). Building on this, results also show that concerns related to information privacy affect individuals’ intentions to use online services (Pavlou et al., 2007; Bélanger et al., 2002), with greater concerns leading to lower intentions. Concerns also lead to individuals being less willing to share personal information with websites (Bélanger et al., 2002).
11 between privacy concerns and the adoption and WTP of such a service. Elaborating on that, one can also assume there is a positive relationship between these concepts when the cloud storage service in question has full privacy protection. This results in the following hypothesis:
H2: Privacy concerns positively affect (moderate) the relationship of the level of privacy protection on the WTP and adoption rate.
For the purpose of measuring privacy concerns both Smith et al. (1996) and Malhotra et al. (2004) developed instruments for testing a person’s privacy concerns. Smith et al. (1996) developed an instrument to measure the concern for information privacy (CFIP), which is composed of 15 items with four subscales tapping into dimensions of individuals’ concerns about organizational information privacy practices. These subscales are: collection, errors, unauthorized secondary use, and improper access. Some years later, Malhotra et al. (2004) adapted and extended the CFIP instrument to explain more of the variance in a person’s willingness to transact, and to make it more applicable to the Internet environment. The Internet user’s information privacy concerns (IUIPC) contains 27 items divided over seven subscales: control, awareness (of privacy practices), collection, errors, unauthorized secondary use, improper access, and global information privacy concern (Malhotra et al., 2004). In choosing one of these constructs for measuring privacy concerns I chose to use the concern for information privacy (CFIP) concept, developed by Smith et al. (1996). Firstly, I chose the CFIP because it is more widely used in the literature and therefore offers more possibilities for comparing my research to other academic work. Secondly, I chose the CFIP to try and keep my survey concise. Where the CFIP instrument has 15 items, the IUIPC has 27, and because this measures only one construct in my research, using the IUIPC would make my survey unnecessarily long, possibly resulting in fewer respondents and therefore less reliable results.
2.5 Socio-demographics
Consumers’ concerns regarding secondary disclosure, as part of information privacy, are likely to be influenced by consumer characteristics. Previous research suggests that socio-demographic characteristics influence online privacy concerns (D’Souza & Phelps, 2009). Explained through the concept of privacy concerns, these socio-demographic characteristics will almost certainly influence the relationship between the level of privacy protection and the WTP and adoption rate of a cloud storage service.
2.5.1 Age
12 often related to opinions about privacy, but the relationships are too complex and inconsistent across studies to reveal a clear pattern. Westin (1991) supports this by pointing out that it is complex how age is affecting privacy attitudes, and that it is possibly being accounted for by generational imprinting, life stage changes and/or experience with institutions.
Despite recognizing the complexity of the effect age has, Westin (1991) came to the conclusion that young people where least concerned about threats to privacy, while older people were much more concerned about threats to their privacy. Supporting this, Sheehan (2002) reported that age is positively related to privacy concerns. Younger consumers were less concerned about privacy issues than those who were older. On top of that, Goldfarb and Tucker (2012) found that older people are much less likely to reveal information than younger people, indicating more privacy concerns. Apart from these peer reviewed articles, many social commentators have expressed the belief that “young people don’t care about privacy”, and this sentiment appears to have caught on in the popular culture (Regan et al., 2013). According to Regan et al. (2013), this sentiment has caught on as a way of interpreting young people’s willingness to post vast quantities of information on social networking sites and to conduct much of their lives in this relatively public space. Hoofnagle et al. (2010) however, had a contradictory conclusion. They concluded that young adults do care about privacy and that young and older adults are actually quite alike on many privacy topics. But since the majority of the studies conclude that age positively affects privacy concerns, this is the assumption I will build on.
H3a: Age positively influences (moderates) the relationship between the level of privacy
protection and the WTP and adoption rate.
2.5.2 Gender
Another socio-demographic characteristic Westin (1991) found to have influence on privacy concerns is gender. In her research on gender differences in online privacy concerns, Sheehan (1999) found that women and men differed significantly in their attitudes toward several practices, with women generally appearing more concerned about the effect some practices would have on their personal privacy. In other studies Dutton and Meadow (1985) and O’Neil (2001) came to the same conclusion. However, Sheehan (1999) also found that men were likely to adopt behaviors to protect their privacy when they became concerned; women, however, rarely adopted protective behaviors. This finding can be of importance when looking at the effect of gender differences on the adoption of the new cloud storage service through the construct of privacy concerns.
H3b: Gender will influence (moderate) the relationship between the level of privacy protection and
13 2.5.3 Household income
The third socio-demographic characteristic that affects privacy concerns is income (Westin, 1991). In combination with privacy concerns, income has not been studied by many researchers. Some used household income as a measure of people’s willingness to share personal information, instead of measuring the effect income has on privacy concerns. However, there are some studies that do investigate this relationship. Dutton and Meadow (1985) found that higher-income respondents are less concerned over privacy invasions. In a more recent study, O’Neil (2001) found the same so be true about the effect of income on online privacy concerns. Although previous literature is quite clear on the effect household income has on privacy concerns, it has to be seen if this can be translated to a lower WTP and adoption rate, since a higher income usually means more purchasing power. Through this logic, the effect that household income has on privacy concerns could be canceled out by a different perception of what the relative value of money is. Although this is something to be reckoned with, in stating the hypothesis I will follow previous academic work and state it as follows:
H3c: Household income negatively influences (moderates) the relationship between the level of
privacy protection and the WTP and adoption rate.
2.5.4 Educational background
As stated earlier, many researchers are convinced that educational background affects privacy concerns (Westin, 1991; Sheehan, 2002; Dutton & Meadow, 1985; Phelps et al., 2000; Zukowski & Brown, 2007). However, there are some contradictory findings. Where some find a positive relationship (Sheehan, 2002), others conclude there is a negative relationship between educational background and privacy concerns (Dutton & Meadow, 1985; Phelps et al., 2000; Zukowski & Brown, 2007). It has to be noted that the study of Meadow and Dutton (1985) specifically focused on how educational background influenced privacy concerns about governmental privacy practices. It will be interesting to see what the effect of educational background is in the specific situation of this study, especially because there are even studies that did not find a difference in privacy concerns between people with different levels of education (O’Neil, 2001). For the purpose of building a hypothesis I will assume that there is a negative relationship, since this is what most studies conclude.
H3d: Educational background negatively influences (moderates) the relationship between the
level of privacy protection and the WTP and adoption rate.
2.6 Mediated moderation
14 The effects of the socio-demographic characteristics (moderators) on the relationship between the level of privacy protection of the cloud storage service and the WTP and adoption rate (main effect), through the concept of privacy concerns (mediator), is assumed to be one of mediated moderation. To understand the concept of mediated moderation it is important to have a clear understanding of what both moderation and mediation represent. In general terms, a moderator is a qualitative (e.g., sex, race, class) or quantitative (e.g., level of reward) variable that affects the direction and/or strength of the relation between an independent or predictor variable and a dependent or criterion variable (Baron & Kenny, 1986). In other words, the overall treatment effect varies in magnitude as a function of the value of the moderator (Muller et al., 2005). In the case of this study, it is hypothesized that socio-demographic characteristics (age, gender, household income, and educational background) moderate the relationship between the level of privacy protection and the WTP and adoption rate of the new cloud storage service.
In general, a given variable may be said to function as a mediator to the extent that it accounts for the relation between the predictor and the criterion. Mediators explain how external physical events take on internal psychological significance. Whereas moderator variables specify when certain effects will hold, mediators speak to how or why such effects occur (Baron & Kenny, 1986). So, the mediator represents the process through which the treatment effect occurs, that is, it explains what process is responsible for the occurrence of a certain treatment effect. In this case privacy concerns are hypothesized as being a mediator between the moderating effects of the socio-demographic characteristics and the main effect.
The effect of the socio-demographic characteristics on the relationship between privacy protection of the cloud storage service and the WTP and adoption, through the concept of privacy concerns, is assumed to be one of mediated moderation. Mediated moderation can happen only when moderation occurs: the magnitude of the overall treatment effect on the outcome depends on the moderator. When the magnitude of the treatment effect depends on an individual difference or context variable, then the mediated moderation question is concerned with the mediating process that is responsible for that moderation (Muller et al., 2005).
15 Concluding, I expect the moderating effects of the socio-demographic characteristics to be at least partly explained by privacy concerns. In that sense, privacy concerns operate as a mediator for the moderating effect that the socio-demographic characteristics have on the main effect (H1). I hypothesize this in the following manner:
H4: The moderating effects that socio-demographic characteristics have on the relationship between the level of privacy protection and the WTP and adoption rate, are at least partly explained (mediated) by the underlying principle of privacy concerns.
2.7 Interaction effects
In this study there are some possible interaction effects that have to be accounted for. Earlier research has shown that there might be an interaction between privacy protection and price. For instance, D’Souza and Phelps (2009) in their study found an interaction effect between price and secondary disclosure, which is the other side of the same coin as privacy protection. Their results show that consumers were more sensitive to price when the company did not share customer information. This sensitivity to price is likely to be driven by the consumers’ perception of secondary disclosure as an additional cost for obtaining the good, or as a hygiene factor. Consumers perceptions of the total cost would accordingly then vary depending on whether the company practices secondary disclosure or not (D’Souza & Phelps, 2009).
Another likely interaction effect that may present itself is the interaction between brand and privacy protection. Closely related to people’s willingness to disclose sensitive information is the degree to which they trust the data gathering entity (Sheehan & Hoy, 2000). Early research, mostly related to the concept of trust, has shown that consumers have a higher willingness to disclose personal information to well-known companies (Earp & Baumer, 2003; Cranor et al., 1999). Elaborating on that, consumers are probably less concerned about privacy protection with regard to services offered by companies they know and trust.
2.8 Characteristics of cloud storage services
16 information, collects customer information for own use, and sells customer information to third parties, where the first refers to the highest level of privacy protection.
For coming up with the levels for the attributes price per month and storage space, I looked at what levels are prevailing in the current market. For price per month that meant levels of €0, €15, and €30. Here, €30 is slightly higher than the price for the most expensive cloud storage service for private use that is on the market at the moment. For storage space the levels are 100 GB, 500 GB, and 1000 GB. These values are without a doubt the most prevailing in the current market.
The better a company’s reputation, the higher its chances are of getting a favorable first hearing for a new product and of getting early adoption of that product (Herbig & Milewicz, 1995). It is therefore very likely that brand has an influence in the case of the adoption of a cloud storage service. Buyers tend to use brand names as signals of quality and value, and often gravitate to products with brand names they have come to associate with quality and value (Herbig & Milewicz, 1995). Resulting from this, the new service will most probably have a disadvantage as opposed to services of established companies like Google and Dropbox. For the conjoint experiment I will use Google, Dropbox, and SugarSync. Google probably is the most well-known brand of the three, followed by Dropbox, which I think is also known by the majority of consumers. SugarSync is the least well-known brand of this selection. Therefore, I think that these three brands can help me account for a company’s reputation and brand recognition.
In their paper, Menard et al. (2014) view the integration of multiple devices as part of a convenience construct that can be seen as a means of minimizing effort (Yale & Venkatesh, 1986). The ability to synchronize data across all devices can be appealing to a user because it eliminates the tediousness of performing the task manually (Menard et al., 2014). Menard et al. (2014) found that adoption of a cloud service for data backup is higher when the users’ perception of integration of multiple devices is high. In my conjoint experiment I will only use the division between full- and no integration of multiple devices for the purpose of keeping the experiment simple and concise.
17
2.9 Conceptual model
The previously stated hypotheses result in the following conceptual model.
18
3. Methodology
In order to collect the data necessary for testing the stated hypotheses I tried to conduct an online survey amongst a representative part of the Dutch population. This survey was spread mainly through Facebook and e-mail. In order to reach respondents that are as diverse as possible I sought assistance of multiple people in different classes of society to spread the survey among people they knew. Nevertheless, the sample drawn from the survey is a sample of convenience.
3.1 Survey design
The survey consisted of three parts. Firstly, questions were asked about socio-demographic characteristics (age, gender, household income, and educational background) and relevant experiences of respondents. Secondly, a choice experiment was conducted. And in the third part respondents answered 15 questions about their privacy concerns.
In the second part, respondents got a series of choice sets with three services that were composed of six attributes (brand, price per month, storage space, level of privacy protection, integration of multiple devices, and quality rating). They had to choose the service they liked best and also answer the question if they would indeed buy this service or not. In total each respondent answered ten of these choice sets. A choice set was visually represented in the following manner:
Service 1 Service 2 Service 3
Brand SugarSync Google Dropbox
Price per month €0 €15 €30
Storage space 100 GB 500 GB 1000 GB
Level of privacy protection Sells customer information to third parties Doesn’t use customer information Collects customer information for own use
Integration of multiple devices Full integration No integration Full integration
Quality rating 4 stars 5 stars 3 stars
[ ] [ ] [ ]
Would you buy the service you have chosen?
[ ] Yes [ ] No
Figure 2: Design of choice experiment
19
3.2 Sample size
To estimate the amount of respondents needed for validly testing the stated hypotheses, I used a measure developed by Johnson & Orme (2003) that calculates the minimum number of respondents needed for a successful Choice-Based Conjoint experiment. The estimate can be calculated by solving the following equation:
𝑛𝑡𝑎
𝑐 ≥ 500
Where n is the number of respondents, t is the number of tasks, a is the number of alternatives per task (not including the no-choice alternative), and c is the number of analysis cells. When considering main effects, c is equal to the largest number of levels for any one attribute. If one is also considering all two-way interactions, c is equal to the largest product of levels of any two attributes (Johnson and Orme, 2003; Orme, 2006). Following this equation, the minimal sample size for this experiment is 150, which looked like a realistic target. However, to be complete, Orme (2006) himself stated that, over the years there became concern that practitioners used this rule-of-thumb to justify sample sizes that were too small. While 500 was set as the bare minimum amount of representations per main effect level in this equation, it would be better to have a 1,000 or more representations per main effect level. In the case of this study it would mean a sample size of about 300 is the minimum.
Unfortunately, I did not manage to get enough respondents when considering this stated minimum. However, the amount of 133 respondents I managed to get should be sufficient to draw some interesting conclusions and obtain valuable insights when analyzing the data.
3.3 Statistical methods
For analyzing and testing the stated hypotheses I used multiple Choice-Based Conjoint analyses (CBC). These models were analyzed separately and, when appropriate, were compared to other models using discriminant analysis.
20
4. Results
Of the 291 started surveys, only 133 where filled in completely. Most of the surveys that where not entirely filled in were stopped right after they started. For the purpose of estimating the model and interpreting the results, only those surveys that where filled in completely are used. Due to the nature of the choice based conjoint data, every respondent has 30 lines of data in the used dataset, one for every possible choice (three per choice set) of all choice sets (ten choice sets per respondent). This means a total of 3990 observations.
The data was checked for missing values and outliers. Here, only one problem arose. In general, the answers given on the questions about the amount of money respondents would be prepared to pay on a yearly and a monthly basis for a cloud storage service that offers full privacy protection are very illogical. A lot of the respondents did fill in the same amount of money for both questions. Some respondents even answered that they would pay a higher monthly fee than total yearly fee. Due to the inconsistency and lack of reliability these variables were not used for further analysis.
4.1 Descriptive statistics
The fact that a convenience sample was used can be seen clearly just by looking at the socio-demographic characteristics, since the values with the highest frequencies all are characteristic to the average student (low Age and Household income). Having said that, when looking at the frequencies of the other possible responses it seems that, although a convenience sample was used, there are enough respondents from other social classes to work with.
Socio-demographic characteristics Percentage of respondents
Gender Male 66.2 Female 33.2 Age 18 – 24 36.8 25 – 34 17.3 35 – 44 7.5 45 – 54 14.3 55 – 64 23.3 65 or older 0.8
Education High school or equivalent 16.5
Bachelor’s degree 47.4
Master’s degree 29.3
Doctoral degree 6.8
Household income Less than €15,000 36.1
€15,000 – €29,999 11.3
€30,000 – €44,999 11.3
€45,000 – €59,999 13.5
€60,000 – €74,999 12.0
More than €75,000 15.8
21 In the process of checking the data for multicollinearity, a correlation matrix was created for these socio-demographics. Since these are mostly ordinal variables, the Spearman correlation is used. Only
Age and Household income show a severe positive correlation (Correlation Coefficient = .774, p =
.000). This correlation is easily explained by the lifecycle an average human being goes through. Older people usually earn more money than younger people due to having a (better) job, especially since the younger respondents are students.
Correlations
Age Gender Education Household
income Spearman's rho Age Correlation Coefficient 1.000 .026 .172** .774** p-value (2-tailed) . .101 .000 .000 Gender Correlation Coefficient .026 1.000 .111** -.025 p-value (2-tailed) .101 . .000 .115 Educational background Correlation Coefficient .172** .111** 1.000 .195** p-value (2-tailed) .000 .000 . .000 Household income Correlation Coefficient .774** -.025 .195** 1.000 p-value (2-tailed) .000 .115 .000 .
**. Correlation is significant at the 0.01 level (2-tailed).
Table 2: Correlation matrix, socio-demographics
To see if this brings about problems when estimating the model, a simple linear regression is performed to get the Variance Inflation Factors (VIF) for all independent variables (IV’s), including the service attributes used for the conjoint analysis. A VIF score for an IV of more than 5 indicates a case of multicollinearity regarding that variable. Highest VIF scores found are scores of 2.522 for
Household income and 2.493 for Age. Both are well below the threshold of 5, so no problems because
of multicollinearity are expected.
4.2 Quality of scale for privacy concerns
22 To test the quality of this scale Principal Axis Factoring is used, while making use of Varimax rotation (when estimating a solution with multiple factors) to make variables load more distinctly on a certain factor. Principal Axis Factoring is chosen because it is more commonly reported in social and behavioral science research reports than, for instance, factoring based on Principal Components. Furthermore, Russell (2002) states that, although many researchers believe there is little difference between Principal Components analysis and Principal Axis Factoring in the results that are obtained, the findings reported by Widaman (1993) indicate that results based on Principal Axis Factoring are more accurate in reproducing the population loadings. Therefore, this appears to be the preferable method of factor extraction (Russel, 2002). First a factor analysis is done with the amount of factors set to one, so that all questions load into one and the same factor.
The correlation matrix of the 15 questions reveals that all questions correlate significantly, some more than others (Appendix 1). Also the Kaiser-Meyer-Olkin Measure of Sampling Adequacy (KMO)(.913), the Bartlett’s Test of Sphericity (p = .000), and the Cronbach’s Alpha (.918) indicate that these questions should factor well. Furthermore, the communalities suggest that all questions should load well on a factor, since, apart from question 13, they all show values of well over the threshold of .4 (Appendix 1).
KMO and Bartlett's Test
Kaiser-Meyer-Olkin Measure of Sampling Adequacy. .913
Bartlett's Test of Sphericity Approx. Chi-Square 34956.818
Degrees of freedom 105
p-value .000
Table 3: KMO and Bartlett’s test Reliability Statistics
Cronbach's Alpha Cronbach's Alpha Based on Standardized Items N of Items
.918 .921 15
Table 4: Cronbach’s Alpha
23 Bearing in mind the initial manner in which Smith et al. (1996) proposed to calculate the total score on their scale, a factor analysis with a four factor solution will now be estimated. With a four factor solution, communalities suggest that all questions should load well on a factor, since they all, apart from question 13 (communality = .338), show values of well over the threshold of .4.
Although the aforementioned statistics tell us that it is highly appropriate to factor the questions concerning privacy concerns, they seem to factor different from what is expected with regard to the works of Smith et al. (1996). Questions about Collection, Errors, and Improper Access load on their own factor as expected. However, the four questions about Unauthorized Secondary Use don’t load as Smith et al. (1996) described. These variables don’t load on the fourth factor, but two of them load on the Collection factor and two load on the Improper Access factor, resulting in a fourth factor in which no variable loads. Full results of the four-factor solution can be found in Appendix 3.
24
Table 5: Three-factor solution, total variance explained Rotated Factor Matrixa
Factor 1 2 3 PC_1_Collection .675 .184 .236 PC_2_Errors .104 .584 .180 PC_3_Un_Sec .364 .145 .590 PC_4_Impr_Acc .252 .355 .601 PC_5_Collection .736 .253 .175 PC_6_Errors .248 .739 .273 PC_7_Un_Sec .437 .355 .287 PC_8_Errors .218 .709 .246 PC_9_Impr_Acc .266 .316 .655 PC_10_Collection .804 .208 .341 PC_11_Un_Sec .615 .200 .378 PC_12_Errors .277 .902 .093 PC_13_Un_Sec .303 .094 .487 PC_14_Impr_Acc .267 .467 .517 PC_15_Collection .641 .228 .327
Extraction Method: Principal Axis Factoring.
Rotation Method: Varimax with Kaiser Normalization. a. Rotation converged in 6 iterations.
Table 6: Three-factor solution, rotated factor matrix
Total Variance Explained
Factor Initial Eigenvalues Extraction Sums of Squared Loadings Rotation Sums of Squared Loadings
Total % of Variance Cumulative % Total % of Variance Cumulative % Total % of Variance Cumulative % 1 7.191 47.940 47.940 6.791 45.276 45.276 3.249 21.662 21.662 2 1.605 10.700 58.640 1.246 8.304 53.580 3.037 20.250 41.912 3 1.035 6.898 65.538 .604 4.030 57.610 2.355 15.698 57.610 4 .747 4.977 70.515 5 .690 4.600 75.115 6 .605 4.033 79.148 7 .560 3.735 82.883 8 .492 3.279 86.162 9 .457 3.050 89.212 10 .347 2.312 91.524 11 .332 2.212 93.736 12 .302 2.011 95.747 13 .240 1.601 97.348 14 .208 1.389 98.737 15 .190 1.263 100.000
25
4.3 Choice-Based Conjoint analysis
For the purpose of performing the basic Choice-Based Conjoint analysis LatentGOLD Choice 5.0 was used.
The model that was estimated first contained only the attributes: Brand, Price per month, Storage
space, Level of privacy protection, Integration of multiple devices, and Quality rating and the
no-choice option. To say something about the goodness of fit, the estimation model is compared to the NULL model and the Chi-square is calculated. In calculating the Chi-square the log-likelihood of the
NULL model (LL(0) = -1328.61) is compared to that of the newly estimated model (LL(β*) = -1001.22). The Chi-square is then being calculated by fitting these into the following equation:
𝐶ℎ𝑖𝑠𝑞 = −2(𝐿𝐿(0) − 𝐿𝐿(𝛽∗))
This results in a Chi-square of 654.7962. After calculating this, the significance of the difference between the two models can be found in a Chi-square table. In this case p(Chisq) < .0001. Since p(Chisq) < .0001, it can be said that the estimated model parameters are significantly different from zero and the model clearly outperforms the NULL model. Furthermore, McFadden’s R² is calculated and supports the finding that the estimation model outperforms the NULL model with an R² of .246 and an adjusted R² of .239. Lastly, the hit rate of the model is calculated and is found to be (299+288+293)/1340 x 100% = 66.17%, which also indicates a good fit of the estimation model.
Estimated /Observed 1 2 3 Total 1 299 71 64 434 2 86 288 79 453 3 68 82 293 443 Total 453 441 436 1330
Table 7: Classification table, basic model
Price per month, Storage space, and Quality rating are presumed to be numeric values, the others are
used as nominal variables. So, where values for the former are depicted as linear, the latter variables are assigned part-worth utilities. These last variables are effect coded. For both Storage space and
Quality rating it was tested if a numerical interpretation was appropriate by comparing the
log-likelihoods and calculating the Chi-squares of different models, with Storage space and Quality rating either used as numeric or nominal variable. The differences in log-likelihood were insignificant for both variables (p(Chisq) > .05). So, to keep working with the more parsimonious model, these attributes were assumed to be numeric.
Thereafter, it was tested if the interaction effects, of which the literature said they could be relevant, were also relevant for this model. After putting in the interaction effects for Brand and Price per
26 terms showed no significance, with p-values of respectively .16 and .83. Therefore, the choice was made to continue with the previous model and disregard the two interaction effects in the rest of the analysis.
Now that it is proven that the estimated model is an appropriate one, we can start interpreting the utilities of the different attributes. Since utilities can only be interpreted relative to the other utilities of the same attribute, and not across attributes, only an indication of their effects can be given. The utilities of all attributes show the relationships one would expect to see. Price per month has a negative utility (U = -1.0378 per level, or U = -.06919 per Euro), indicating that people prefer cheap over expensive services. Storage space has a small positive utility (U = .089 per level), indicating that respondents preferred more over less storage space, but did only attach little value to it. Also, the
Integration of multiple devices is preferred (Uno = -.1871, Uyes = .1871). Furthermore, Quality rating
has a positive utility of .072, indicating people value a higher over a lower rating. The no-choice option is highly insignificant (p = .89). Due to this, the importance of the other attributes might be overstated. Because of this, one might have to interpret the results a little differently to still come to valuable insights. Interpreting the results as “the importance when buying”, instead of “the importance for buying” could be a solution.
For Brand, the well-known brands (UGoogle = .0897, UDropbox = .2176) are clearly preferred over the lesser known brand (USugarSync = -.3072). It was expected that Google would be the most preferred brand, instead results show that Dropbox is the most preferred one. However, when looking deeper into the data, this fully supports the expectation that the best known brand is preferred. When filling in the survey, 72 out of the 104 respondents that have used cloud storage services before indicated to have used Google Drive before. When compared to the 100 respondents said to have used Dropbox before, it becomes apparent that not Google, but Dropbox is the most well-known cloud storage service.
Lastly, the attribute that is the most important for this study, Privacy protection, shows the expected utilities. The highest level of protection has the highest utility (U = .6232) and the lowest level of protection has the lowest utility (U = -.7189). Since a positive utility function is direct proof that a higher Level of privacy protection leads to a higher adoption rate, and that consumers’ WTP increases when Level of privacy protection increases, we can already say that H1 is supported.
H1: When the level of privacy protection of the cloud storage service gets higher, the WTP and adoption rate will get higher. Supported
27 by Level of privacy protection (28.9%), Brand (11.3%), Integration of multiple devices (8.0%),
Storage space (3.8%), and Quality rating (3.1%). It is interesting to see that Level of privacy protection is deemed to be the second most important attribute for respondents when choosing
between different cloud storage services. Apparently people are genuinely concerned about their privacy when it comes to cloud storage services. All results can be found in table 8.
To see how much money people are prepared to pay for a cloud storage service that offers full privacy protection, the WTP is calculated by dividing the utilities of the different levels of privacy protection by the utility for Price per month per Euro. Due to the nature of the data and the calculation of WTP a negative outcome represents a positive WTP. After setting the level that offers no privacy protection as reference, it is estimated that people are willing to pay €19.40 per month for full privacy protection compared to no privacy protection. This seems to be rather high in relation to the research done by Ion et al. (2011), where they found that 79 percent of consumers was willing to pay 20 USD per year for retaining full privacy.
In light of the concepts of willingness-to-protect and willingness-to-sell, developed by Grossklags and Acquisti (2007), it might be more realistic not to calculate the WTP with the lowest Level of privacy
protection set to zero, but to calculate the WTP as the utilities are, i.e. take into account both the
negative WTP (willingness-to-sell) and the positive WTP (willingness-to-protect). When interpreting the results in this manner a WTP of only €9.01 is calculated. At this point it is difficult to judge which of these values comes closer to the truth. Because of this, I will wait with the definitive interpretation of these results. Results can be found in table 9.
Attributes Class1 Wald p-value Mean Std.Dev. Range Importance Rank Brand
Google 0.0897 40.1269 1.90E-09 0.0897 0 0.5248 11.3% 3rd
Dropbox 0.2176 0.2176 0
SugarSync -0.3072 -0.3072 0
Price per month
-1.0378 483.2499 4.20E-107 -1.0378 0 2.0756 44.8% 1st
Per € -0.06919
Storage space
0.089 4.7266 0.03 0.089 0 0.178 3.8% 5th
Level of privacy protection
1 0.6232 210.74 1.70E-46 0.6232 0 1.3421 28.9% 2nd
2 0.0957 0.0957 0
28
Integration of multiple devices
No -0.1855 26.7202 2.40E-07 -0.1855 0 0.371 8.0% 4th Yes 0.1855 0.1855 0 Quality rating 0.0727 3.1131 0.078* 0.0727 0 0.1454 3.1% 6th No-choice 0 4.2431 0.0179 0.89** 0.0411 0 1 -4.2431
*=Not significant at 0.05, but significant at 0.10 **=Not significant at 0.10
4.6716 100% Total
Table 8: Choice-Based Conjoint analysis, basic model
Level of privacy protection Utility Relative WTP Level 3 set to 0
1 .6232 -€9.01 -€19.40
2 .0957 -€1.38 -€11.77
3 -.7189 €10.39 -
Table 9: WTP for levels of privacy protection
4.3.1 Segmentation of respondents
After all these measures were calculated for the overall model, a latent class analysis is done to see what group of people generally values privacy protection the most and how much money this group is prepared to pay for full privacy protection. After adding Age, Household income, Gender, Educational
background, and Privacy concerns, the model is estimated for one to six classes to come to the
appropriate amount of classes. The model showing the lowest value for the Bayesian Information Criteria (BIC) is the preferred model and will be used for further analysis. After conducting the analysis the five-class model is chosen for having the lowest BIC value (see table 10). The resulting five-class model can be found in table 12.
LL BIC(LL) Npar df p-value Class.Err R²(0) R² 1-Class -1001.216 2046.444 9 124 3.40E-33 0 0.3375 0.3375 2-Class -828.385 1823.043 34 99 7.20E-281 0.0135 0.5009 0.5008 3-Class -742.247 1773.024 59 74 2.70E-261 0.0125 0.5746 0.5746 4-Class -659.940 1730.668 84 49 3.70E-244 0.0155 0.6509 0.6508 5-Class -590.299 1713.647 109 24 3.40E-234 0.0195 0.6953 0.6952 6-Class -546.840 1748.987 134 -1 0.0172 0.7320 0.7320
Table 10: Latent class analysis, results
29 consumers that went for full privacy protection. Furthermore, for class three the Level of privacy
protection is also the most important attribute (44%). The importance of the attributes for each of the
five classes can be found in table 11.
Attribute Importance 1 Importance 2 Importance 3 Importance 4 Importance 5
Brand 7.0% (4th) 22.3% (2nd) 5.9% (4th) 3.1% (5th) 29.5% (1st)
Price per month 69.5% (1st) 24.4% (1st) 33.5% (2nd) 20.4% (2nd) 24.4% (3rd)
Storage space 7.3% (3rd) 5.8% (6th) 13.2% (3rd) 14.2% (3rd) 13.9% (4th)
Level of privacy protection 7.9% (2nd) 17.8% (4th) 44.0% (1st) 58.5% (1st) 1.6% (6th)
Integration of multiple devices 5.9% (5th) 21.1% (3rd) 1.2% (6th) 3.6% (4th) 2.6% (5th)
Quality rating 2.4% (6th) 8.7% (5th) 2.2% (5th) 0.3% (6th) 28.0% (2nd)
Table 11: Choice-Based Conjoint analysis, importance of attributes, five-class model
Table 12: Choice-Based Conjoint analysis, attribute measures, five-class model
The measures concerning good model fit seem promising. Compared to the one-class model the R² and R²(0) improved from respectively .3375 and .3375 for the classless model to .6953 and .6952 for the model with five classes, indicating that more of the variance is explained by the new model. Also, according to the Chi-square statistics (Chisq = 821.8332, p(Chisq) < .0001) the model fit improved significantly. Furthermore, both the hit rates for predicting choices (85.71%) and for predicting class
Attributes Class1 Class2 Class3 Class4 Class5 Wald p-value Wald(=) p-value Mean Std.Dev.
Brand
Google 0.0694 0.6335 -0.3267 -0.1526 -0.8747 92.0877 2.10E-15 67.2759 1.70E-15 0.0211 0.4177
Dropbox 0.707 -0.0471 0.2863 0.2081 2.8232 0.5105 0.6913
SugarSync -0.7763 -0.5864 0.0404 -0.0555 -1.9484 -0.5316 0.5114
Price per month
-4.2205 -0.6675 -1.7266 -0.8722 0.4442 160.5288 7.60E-33 76.0804 1.20E-15 -2.0328 1.6376
Per € -0.2814 -0.0445 -0.1151 -0.0585 0.0296
Storage space
-0.4464 0.1585 0.6817 -0.6086 0.2531 32.3903 5.00E-06 28.4626 1.00E-05 -0.0425 0.4782
Level of privacy protection
1 0.2629 0.4435 1.5114 2.9665 -0.0360 151.0815 2.20E-27 92.4191 1.50E-16 0.9547 0.9816
2 0.4286 0.0856 1.5075 -0.9089 0.0133 0.3447 0.7394
3 -0.6915 -0.5291 -3.0189 -2.0577 0.0277 -1.2993 1.0451
Integration of multiple devices
No 0.3597 -0.5769 0.0601 -0.1546 -0.0477 50.6411 1.00E-09 25.5407 3.90E-05 -0.03 0.3542 Yes -0.3597 0.5769 -0.0601 0.1546 0.0477 0.03 0.3542 Quality rating -0.1477 0.2372 0.1118 0.0131 0.5102 9.776 0.082* 4.4498 0.35** 0.0672 0.1913 No-choice 0 2.533 2.6667 2.6404 3.5663 -2.0683 0.0067 1.00** 0.0029 1.00** 2.4268 1.2688 1 -2.533 -2.6667 -2.6404 -3.5663 2.0683 -2.4268 1.2688
30 membership (96.05%) show a good model fit. Classification tables can be found in table 14 and table 15
Table 13: Choice-Based Conjoint analysis, class profile attributes, five-class model
Estimated /Observed 1 2 3 Total 1 369 32 33 434 2 29 391 33 453 3 27 36 380 443 Total 425 459 446 1330
Table 14: Classification table, choices, five-class model Attributes Class1 Class2 Class3 Class4 Class5 Class Size 0.3342 0.2363 0.2104 0.1502 0.0687
Brand
Google 0.3011 0.5551 0.2331 0.2828 0.0240
Dropbox 0.5697 0.2810 0.4303 0.4056 0.9678
SugarSync 0.1292 0.1639 0.3365 0.3116 0.0082
Price per month
€0 0.9853 0.5630 0.8268 0.6293 0.2004 €15 0.0145 0.2888 0.1471 0.2618 0.3124 €30 0.0002 0.1482 0.0262 0.1089 0.4872 Mean 1.0149 1.5851 1.1994 1.4795 2.2868 Storage space 100GB 0.4879 0.2821 0.1452 0.5434 0.2533 500GB 0.3122 0.3306 0.2871 0.2957 0.3263 1000GB 0.1998 0.3873 0.5677 0.1609 0.4203 Mean 1.7119 2.1053 2.4225 1.6175 2.167
Level of privacy protection
1 0.3898 0.4814 0.4983 0.9734 0.3214
2 0.4601 0.3366 0.4964 0.0202 0.3377
3 0.1501 0.1820 0.0054 0.0064 0.3409
Integration of multiple devices
31 Proportional
/Latent Class 1 Class 2 Class 3 Class 4 Class 5 Total
Class 1 43.4666 .6967 .416 .0002 0 44.5795 Class 2 .6967 29.3828 1.2704 .113 .0006 31.4636 Class 3 .416 1.2704 26.1754 .1276 0 27.9895 Class 4 .0002 .113 .1276 19.7262 0 19.9671 Class 5 0 .0006 0 0 8.9997 9.0004 Total 44.5795 31.4636 27.9895 19.9671 9.0004 133
Table 15: Classification table, class membership, five-class model
4.4 Testing the hypotheses
After proving the segmented model is an appropriate one, the other hypotheses will be tested one by one.
4.4.1 Privacy concerns
When comparing class four to other classes on the basis of Privacy concerns it is clear that respondents in this class score relatively high on Privacy concerns. In fact, the average score of the fourth class is 5.3282, the highest of all classes (see table 16). This is a logical result considering this class puts a lot of value on the Level of privacy protection of the service. When looking at the utility of
Privacy concerns in the different classes it seems that these fully support hypothesis H2. The classes
that value the Level of privacy protection the most, namely classes three and four, show positive utilities for Privacy concerns (Class 3: U = 1.0848, Class 4: U = 1.9344), while the utilities in other classes show negative values (p = .0054) (see table 17). This means that consumers that value privacy protection more than others show significantly more privacy concerns than others. In other words, the effect of the Level of privacy protection on the WTP and adoption rate of a cloud storage service is strengthened by the score a respondent has on Privacy concerns, where people with a high score on
Privacy concerns value the Level of privacy protection more. Hypothesis H2 is thus supported.
32
Table 16: Choice-Based Conjoint analysis, class profile covariates, five-class model Covariates Class1 Class2 Class3 Class4 Class5 Age 18 – 24 0.3094 0.5000 0.5526 0.1506 0.1111 25 – 34 0.2460 0.1245 0.0735 0.2033 0.2222 35 – 44 0.1102 0.0640 0.3830 0 0.2222 45 – 54 0.1254 0 0.1960 0.3468 0.1111 55 – 64 0.2090 0.3115 0.1039 0.2993 0.3334 65 or older 0 0 0.0357 0 0 Gender Men 0.7898 0.6918 0.5363 0.4515 0.7778 Women 0.2102 0.3082 0.4637 0.5485 0.2222 Educational background
High school or equivalent 0.1334 0.2061 0.2016 0.0463 0.3334
Bachelor’s degree 0.4477 0.4001 0.5506 0.553 0.4444 Master’s degree 0.3557 0.2607 0.2478 0.3008 0.2222 Doctoral degree 0.0632 0.1331 0 0.0998 0 Household income Less than €15,000 0.3800 0.3764 0.5059 0.2533 0 €15,000 – €29,999 0.0844 0.2608 0.0370 0.0500 0.1111 €30,000 – €44,999 0.1564 0.1475 0.0848 0.0507 0 €45,000 – €59,999 0.0443 0.0638 0.0752 0.3461 0.5556 €60,000 –€74,999 0.1561 0.0390 0.1009 0.1998 0.1111 More than €75,000 0.1787 0.1125 0.1963 0.1002 0.2222 Privacy concerns 1-24 0.3135 0.2029 0.0585 0 0.4444 25-42 0.2168 0.3435 0.1618 0.1002 0.0001 43-57 0.2582 0.1178 0.2427 0.1999 0.1111 58-75 0.0792 0.2062 0.3176 0.3554 0.2222 76-85 0.1324 0.1296 0.2194 0.3445 0.2222 Mean 4.4509 4.5721 5.0206 5.3282 4.3981
33
Table 17: Choice-Based Conjoint analysis, covariate measures, five-class model
4.4.2 Age
Regarding the moderating effect of Age, no clear conclusion can be drawn from the utilities, since no clear pattern can be recognized. On top of that, the results are not significant (p = .47). However, some conclusions can be drawn when looking at the shares of the different age groups in class 4. What is interesting to see is that respondents belonging to this class are relatively old. The percentage of people that is 45 or older in the total sample is 38.4%. This same group makes up 64.61% of the fourth class, which is very sensitive for privacy protection. This difference seems to be significantly big enough to draw conclusions from. Concluding from this, it can be said that older people tend to value privacy protection in a cloud storage service more than young people.
H3a: Age will positively influence (moderate) the relationship between the level of privacy
protection and the WTP and adoption rate. Supported
4.4.3 Gender
In the earlier estimated model with latent classes, Gender proved to be heterogeneous between classes (p = .048), where women had a significantly bigger chance of belonging in class 4 than men (UWomen = .7041 vs UMen = -.7041). Also, when looking at the percentage of women in the privacy protection sensitive class (54.85%) compared to the percentage of women in the total sample (33.2%), there is a significant difference between the two. Therefore, H3b is supported.
H3b: Gender will influence (moderate) the relationship between the level of privacy protection and
the WTP and adoption rate, in that women will have a higher WTP and adoption rate than men. Supported
Educational background
High school or equivalent -0.3276 -0.1726 1.7053 -2.5173 1.3122 10.0283 0.61**
Bachelor’s degree -1.0108 -1.1483 0.7648 0.8650 0.5293 Master’s degree -0.4151 -0.9877 0.6483 0.6421 0.1118 Doctoral degree 1.7529 2.3087 -3.1184 1.0101 -1.9532 Household income Less than €15,000 0.6362 0.8746 0.7320 0.5530 -2.7957 26.4953 0.15** €15,000 – €29,999 0.2144 1.6323 -1.2508 -1.5408 0.9448 €30,000 – €44,999 0.9385 2.0432 0.0828 -0.9375 -2.127 €45,000 – €59,999 -2.4940 -1.4519 -0.2726 2.4671 1.7513 €60,000 –€74,999 0.5214 -1.9381 0.2844 0.5403 0.5920 More than €75,000 0.1834 -1.1602 0.4242 -1.0821 1.6347 Privacy concerns -1.1404 -0.8069 1.0848 1.9344 -1.0719 14.6868 0.0054
