The effect of Heart Disease Complaints, Smartwatch Brand Trust
and Reputation of Data Receiver on Willingness to Disclose
Information Through Smartwatches
By: M.F. Kuipers Student number: 11413743
MSc. in Business Administration – Digital Business Track University of Amsterdam
Supervisor: Prof. em. dr. ir. H.J. Oppelland Date: June 22, 2018
Statement of originality
This document is written by Myrthe Kuipers 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
Contents
Abstract ... i Introduction ... 1 1.1 Research problem ... 2 1.2 Research objectives ... 4 1.3 Method ... 4 1.4 Structure ... 5 2. Literature review ... 5 3. Research design ... 153.1 Sample and response ... 16
3.2 Survey design and measures ... 17
4. Results ... 20
4.1 Statistical procedure ... 20
4.2 Analyses outcomes ... 25
5. Discussion and implications ... 36
6. Limitations and research agenda ... 42
Bibliography ... 46
Appendices ... 52
Appendix 1 Notes meeting Ellen Janssens Dutch Heart Foundation ... 52
Appendix 2 Notes meeting Gregory M. Marcus (University of California San Francisco)*54 Appendix 3 Variables including validated measures... 56
Appendix 4 Hypotheses ... 57
Appendix 5 Survey (in Dutch) ... 58
List of figures and tables
Figure 3.1 Conceptual framework p.15
Table 3.1 Survey Scenarios p.17
Table 4.1 Descriptive Statistics p.21
Table 4.2 Reliability of scale items p.23
Table 4.3 Means, Standard Deviations and Correlations p.25
Table 4.4 Linear Regression CVD, SBT and RDR on PB p.28
Table 4.5 Logistic regression models predicting ID p.28
Table 4.6 Results mediation of PB between CVD and ID p.30
Table 4.7 Mann-Whitney U results for mean differences in ID by CVD groups p.30
Table 4.8 Results mediation of PB between SBT and ID p.32
Table 4.9 One-way ANOVA test for SBT p.33
Table 4.10 Mann-Whitney U results for mean differences in ID by SBT groups p.33
Table 4.11 Results mediation of PB between RDR and ID p.35
Table 4.12 One-way ANOVA test for RDR p.35
i
Abstract
E-health not only enables healthcare professionals to improve treatment performances and communication with patients, but also enables new forms of data collection through devices such as smartwatches. This study examines the effect of cardiovascular disease (CVD) patients and non-patients, the reputation of data receivers and smartwatch brand trust on someone’s willingness to disclose personal health information to healthcare professionals. Perceived benefits and perceived coping capabilities of sharing information were explored for their mediating role in this process. The aim of this study is to provide healthcare professionals implications on (1) how to approach patients and non-patients for data collection, (2) what type of smartwatch brand to consider while collecting data and (3) whom to share data with. A survey was filled out by 253 respondents who were exposed to different type of scenarios in which they had to evaluate their perceived benefits and perceived coping capabilities based on the reputation and trust score they assigned to a specific data receiver (i.e. Erasmus Medical Center or Menzis health insurance) or smartwatch brand (i.e. Apple or Zeblaze). Results indicate that CVD patients are more willing to share data than non-patients and that the reputation of the data receiver has more effect on information disclose than a smartwatch brand. Therefore healthcare professionals should prioritize improving their reputation and then consider familiar over unfamiliar smartwatch brands to improve individuals’ willingness to disclose information.
Keywords: E-health, smartwatch, brand trust, privacy awareness, privacy benefits, privacy concerns, heartrate measurements, information disclosure.
1
Introduction
E-health has become a popular topic over time. Especially with the rise of new technology, such as smartwatches and real-time tracking mobile applications over the past few years. This rise has potential to result in improving health outcomes, like for instance more accurate measurements, less errors in prescriptions and the ability for individuals to have access to their own health records and performances (Angst, 2009; Beranek Lafky & Horan, 2011). However, patients worried about privacy may avoid such technologies which require patients to share personal information (Lohr, 2015). The author states that not every individual may be willing to share personal data, but potential benefits (e.g. understanding treatments and its performance and remove errors) can outweigh risks.
Healthcare professionals and in this particular research scope heart disease professionals, are increasingly investing in new technologies to measure personal health records to improve healthcare (appendix 2). For instance, due to the availability of accurate insights in patients’ data cardiologists will be able to spend time more efficiently with their patients (O’neil, 2017). Parties such as the Dutch Heart Foundation, cardiologists and hospitals desire to collect data (e.g. heartrates and sociodemographic characteristics) of individuals to gain more insights in causes of heart diseases and to use these insights to heal and prevent them. As such, researchers are working on methods to detect what is causing atrial fibrillation and use these data in algorithms. Gregory M. Marcus (cardiologist and electrophysiologist practicing at UCSF) is one of these researchers: “It is all just about using heartrate data, in which watches can help to measure.” (appendix 2). Additionally, Marcus and his research team claimed that Apple watches are now able to detect atrial fibrillation by 97% accuracy during the Heart Rhythm Society conference of 2017 in Chicago (Sanchez, 2017). Smartwatches are therefore interesting instruments for health researchers to consider while tracking patients’ health.
2 Furthermore, health insurance companies are interested in receiving such personal health data as well (appendix 1). Health insurance companies may use data to separate healthy people from the unhealthy people within a population (O’neil, 2017). While doing so, the insurance companies may be able to predict what kind of health insurance people need. This will enable health insurance companies to adjust services more accurately to the market. In the US healthcare data grew by 48% in 2017 and this data can all be used as input to predict behavior upon (Meeker, 2018). The Dutch Heart Foundation’s Innovation Manager Ellen Janssens stated, “In the past we focused on the healing process of heart diseases. However, considering the current technological possibilities our focus is shifting towards heart disease recognition.” (appendix 1). As a result, the Dutch Heart Foundation collaborated with Philips to work on new technology to measure heartrates (appendix 1). Currently, most data is being collected from patients already suffering from cardiovascular diseases (CVD), which makes tracking risk factors that cause these heart diseases challenging. Therefore, measuring heartrate performances and developments of non-patients is interesting to gain new insights in this particular problem. Considering the rise of E-health, smartwatches can support healthcare professionals in tracking and collecting data from patients as well as non-patients.
1.1 Research problem
Decisions on how to stimulate individuals to share data and which smartwatch brand to consider for data collection has to be made by healthcare professionals in order to successfully obtain data from healthy people and people suffering from CVD complaints. Accessibility and transferability of health data may results in interesting insights for healthcare professionals. However, it is not clear whether patients and non-patients agree on transparency of their data among multiple organizations or if smartwatch brands play a role in willingness to share information. Insights in how trust of smartwatch brands and reputation of information receivers may vary can be used as input within this decision. Previous researches
3 (e.g. Moorman, Christine ; Zaltman, Gerald ; Deshpande, 1992; Wottrich, Verlegh, & Smit, 2017; Xie, Teo, & Wan, 2006) have focused on the influence of trust and reputation on information disclosure, yet the combination of two instances within this decision process has not been examined so far. For instance, are people willing to disclose more personal information to an organization with a weak to moderate reputation when a highly trusted smartwatch is being used as an instrument to measure data with?
Although many researches have been done on privacy behavior and motivations of consumers to disclose personal information (e.g. Acquisti, Brandimarte, & Loewenstein, 2015; Beranek Lafky & Horan, 2011; Ginosar & Ariel, 2017), the effect of someone’s CVD complaints in combination with brand trust of smartwatch devices and the reputation of the data receiver on personal information disclosure through smartwatches has not been studied yet. Beranek Lafky & Horan, (2011) found that privacy concerns do not vary much between healthy and unhealthy people, but that unhealthy people were more tended to share their personal health information. Nevertheless, it is not clear with what types of parties within the healthcare they are willing to share data with. Although the influence of context on willingness to disclose personal information is researched by Bansal, Zahedi, & Gefen (2016), it is unknown whether these contexts are similar to an individual’s perception of reputation of a specific information receiver within the field of healthcare. Brand trust of devices may also play a key role in willingness to disclose information besides the reputation of the party who will receive the data. From the qualitative study of Udoh and Alkharashi (2017) it seems like the brand of a smartwatch influences privacy concerns of US citizens. However, it remains unclear whether this is truly the case and whether results are generalizable to the Dutch population. Therefore, this topic needs further investigation. Implications on which smartwatch brand to consider and how this will affect an individual’s willingness to disclose personal information is necessary since CVD healthcare professionals are considering various
4 smartwatch brands to collect data with (appendix 1). Moreover, healthcare professionals may not be aware whether a smartwatch brand affects someone’s willingness to disclose information yet.
Current literature on privacy behavior does not provide enough input for CVD healthcare professionals to make grounded decisions on what kind of smartwatches to use while approaching healthy people and people with CVD complaints for personal data and whom to share these data with.
1.2 Research objectives
Therefore, the objective of this study is to determine the effects of CVD complaints, smartwatch brand trust and reputation of information receiver on personal information disclosure to healthcare professionals in the fields of heart disease. This research will provide CVD healthcare professionals guidelines on how to make certain decisions on (1) how to approach patients and non-patients for data collection, (2) what type of smartwatch brand to consider to collect data with and (3) whom to share data with. The leading research question of this study is formulated as following: “What are the effects of CVD complaints, smartwatch brand trust and reputation of the data receiver on information disclosure and what role do perceived benefits and perceived coping capabilities have in this relationship?”
1.3 Method
The nature of this research is explanatory to test the hypotheses of this study. Therefore a quantitative survey design is used to measure relationships between variables. In total 253 respondents completed the survey who were collected through availability sampling. Also, an interview has been conducted with an expert in the field of cardiovascular diseases (CVDs) to finalize variables that are included in this study. Additionally, a pre-test was conducted to ensure that survey questions are applicable to and understandable by all respondents and to
5 check whether manipulation within the variables ‘smartwatch brand trust’ and ‘reputation of data receiver’ works. After receiving feedback from eight respondents some grammar mistakes were corrected and descriptions were formulated more clearly.
1.4 Structure
First of all a review of the literature on privacy behavior is provided in order to obtain knowledge which forms the groundwork of this study’s framework and hypotheses. Furthermore, chapter three will discuss the conceptual framework including the design of this study followed by the results in chapter four. Finally, a discussion will be provided along with the limitations of this study which are presented in chapter five and six.
2. Literature review
This study builds upon research of Park & Chung (2017) in which health privacy is being perceived as sociotechnical capital. In their research health privacy capital (HPC) consists of privacy awareness (i.e. level of privacy knowledge which is an enabler of protective behavior and willingness to engage with information technologies), perceived benefits that come from information disclosure (i.e. the extent to which the benefits that come from data sharing outweigh perceived risks) and coping capability of potential risk factors (i.e. one’s ability to have control over personal data and protect themselves from unwarranted surveillance). These three elements are similar to the so called ‘privacy calculus’ which will be discussed further in this literature review.
Park & Chung (2017) stress that one can only effectively use medical technologies once their health privacy capital is sufficient on all three levels. For instance, someone who is quite aware of privacy, but lacks confidence to cope with potential risk factors, may decide to not disclose personal health information. As mentioned earlier, Lohr (2015) stressed that this may result in a barrier within the development to improve healthcare. A high level of health
6 privacy capital enables consumers to actively and sufficiently use digital health resources, which in turn leads to positive health outcomes since data can be shared among multiple parties to improve health. These people tend to seek engagement in digital media more and adapt health technologies faster than patients who are low on health privacy capital. Their study has provided insights in how sociodemographic factors influence HPC within the United States among people within a cancer community. It remains unclear whether their results are generalizable to the Dutch population even though the authors claim their results are generalizable to the US population. Besides, it is not known if these results apply to patients suffering from other diseases rather than solely cancer.
Within the literature of privacy behavior, three levels of analyses have been mapped by Greenaway and Chan (2005), which are individual privacy, public/national privacy and organization oriented privacy. Since this research is mainly concerned with individual’s privacy and health information disclosure, the organizational and national perspectives have been filtered out of scope.
Privacy paradox
The ‘privacy paradox’ has been described as: “the gap between users' concerns regarding the risks of violating their privacy on the one hand, and their willingness to disclose personal information on the other hand.” (Ginosar & Ariel, 2017, p. 949). It is the degree to which consumers (1) are aware of potential benefits that can emerge from sharing information, (2) trust websites or activities on the Internet and (3) are aware of technological characteristics and its risks (also known as ‘online literacy’) (Ginosar & Ariel, 2017). Since privacy awareness plays a critical role within the privacy calculus this concept will be explained first followed by perceived privacy concerns and benefits (i.e. the main privacy calculus) and coping capabilities (i.e. addition to the privacy calculus).
7
Privacy awareness
Being aware of privacy can cause individuals to become protective over their data. In the privacy behavior literature, privacy awareness refers to the extent an individual is aware, or has the knowledge, of what data is being collected by the particular instance (Park & Chung, 2017). Privacy awareness can also be described as ‘online literacy’ and indicates an individual’s level of knowledge of Internet technologies and risks related to that (Ginosar & Ariel, 2017). Bartsch and Dienlin (2016) found that consumers with high online privacy literacy are more cautious when it comes to sharing personal information. The more people become familiar with online activities, the more people become privacy cautious. Park (2013) elaborates on this finding as well as S. Trepte et al. (2014) who found that there are a few dimensions of privacy literacy, namely: (1) knowledge about the activities of the organization in question, (2) knowledge of technical aspects of online privacy and data protection, (3) knowledge of online data protection laws, and (4) knowledge of personal strategies to maintain privacy regulation.
In order for consumers to become aware of privacy, they have to read organizations’ policies. Privacy policies and information on organizations’ websites generally provide individuals insight on whether they have the ability to cope with the risks or not. Studies showed that about half of Internet users read privacy policies (Ginosar & Ariel, 2017). In their literature review they describe how consumers’ willingness to read the conditions is being influenced by language proficiency and the length of the privacy policies. Park (2013) adds that the level of privacy literacy affects consumers’ trust in online activities. Even though individuals are aware of risks related to information disclosure, research by Acquisti, Brandimarte, & Loewenstein (2015) showed that many of these individuals do not take further measures. Also Ellen Janssens admitted she is not sure whether patients are really concerned with privacy risks since she does not receive many questions or comments related
8 to this issue: “I never really receive many complaints or questions on this topic from our community.” (appendix 1).
Perceived risks
Trust has been an important factor within the field of privacy behavior and someone’s willingness to disclose personal information. The literature review by Ginosar & Ariel (2017) pointed out that consumers are mainly concerned with: (1) collection of personal information without informed consent, (2) the degree to how much information is shared with third parties and (3) to which extent information is being used for secondary purposes without the consumer knowing while browsing the Internet. Other privacy concerns found related to wearable devices are the pervasive collection of personal and health data through continuous tracking and opportunistic behaviors that might appear because of this (Marakhimov & Joo, 2017). Furthermore, Beranek Lafky & Horan (2011) found that respondents were mostly concerned with the security of banking information, followed by security of tax data and health data on a shared second ranking. Similarly, Udoh & Alkharashi (2017) showed that respondents were mainly concerned with financial and health information, but most of all with identity theft. People tend to proceed into protective actions, such as refraining from information disclosure and falsifying information when such privacy concerns occur (Matt & Peckelsen, 2016). The literature points out that privacy concerns negatively influence the willingness to disclose information. Nevertheless, Internet device users increasingly accept Health Information Exchange (HIE), although they gain knowledge on privacy and become more privacy cautious over time which has been confirmed by multiple researchers (Acquisti, Brandimarte, & Loewenstein 2015; Udoh & Alkharashi 2017). This finding, which lies in contrast to the results of Matt & Peckelsen (2016), Park (2013) and Bartsch and Dienlin (2016), suggests privacy concerns may play a role in the back of the minds of individuals, but it does not directly affect behavior. A possible reason for this effect are the perceived benefits
9 which may result from disclosure of personal data. This may also be the reason why this topic is not heavily discussed by patients which Ellen Janssens mentioned.
Perceived benefits
During information exchange, consumers outweigh potential benefits to perceived risks of this exchange. In the privacy behavior literature this is also known as the ‘privacy calculus’ (Culnan & Bies, 2003). According to Angst (2009) perceived benefits can either be out of self-interest or out of interest for society. “There appears to be a delicate balance between acting in self-interest (some might argue that self-preservation is a more accurate depiction) or acting in the interest of society as a whole” (p. 175). This indicates that not solely patients may perceive personal benefits (e.g. insights in their own heart health) of disclosing personal data, but also non-patients may perceive benefits (e.g. contribution to society’s health).
Chen (2001) and Rainie and Duggon (2016) found that people are willing to disclose personal information in return for tangible benefits. Tangible benefits in case of sharing personal health information are for instance improving communication between individuals and healthcare providers in question or gaining direct insights in personal health improvements. Nevertheless, societal benefits also play a role in case of cardiologists, among other researchers, desiring to improve research on cardiovascular diseases. Improving healthcare within the Netherlands will be beneficial to the whole Dutch population. These findings indicate that an increase in perceived benefits can result in an increase of willingness to disclose personal health information.
Perceived coping capabilities
Literature on the privacy calculus has been enriched over the past few years and Li (2012) found that a ‘dual-privacy-calculus’ exists. Within this calculus a consumer’s coping mechanism to handle concerns is brought into perspective besides perceived privacy risks and
10 benefits. Coping capabilities refer to the amount of one’s perceived data protection possibilities. Park & Chung (2017) used the term ‘confidence’ which in this study will be treated as a synonym of ‘perceived coping capabilities’. They describe this term as the possibility of an individual to have a say in what data is allowed to share and not to share and whether an individual feels confident in the data protection of technology. Li's (2012) research pointed out that consumers’ ability to cope with potential risks positively influences personal information disclosure. Threat appraisal or coping determines individuals’ protection behavior, which in health research is the most immediate predictor of behavior (Webb, 2006). Individuals are less likely to disclose personal information in cases of insufficient coping capabilities. Hence, an increase in perceived coping possibility is expected to result in an increase in willingness to disclose personal health information.
The privacy calculus is well known within the privacy behavior literature and has been explained in general. In this study the privacy calculus will remain central and in the next section of the literature review health (CVD) complaints, smartwatch brand recognition and type of data receiver will be introduced in relation to the privacy calculus in order to achieve this study’s objective.
CVD complaints
In context of this research, heart attacks, stroke, heart failure and rheumatic heart disease are forms of cardiovascular diseases (CVDs). Not much research has been done in relation to healthy against unhealthy people and their attitudes towards privacy and willingness to disclose personal health information. However, Beranek Lafky and Horan (2011) studied attitudes towards health information privacy and interviewed 17 respondents from different health statuses (i.e. well, unwell and disabled respondents). While using the Delphi technique, they found that health information privacy attitudes between well and unwell people do not vary much. Nevertheless, they also conducted a survey among 230 respondents and found that
11 people who are well have more privacy concerns compared to people who are disabled or who have a chronic illness. Also, the analyses showed that unwell and disabled people were most likely to share their health information with others and results show that people do not state to be overly concerned with their health privacy. Notably, disabled and unwell people tend to be less secure when it comes to data protection compared to healthy people. It must be noted that in their research mostly covered personal benefits of health information disclosure (e.g. access to the correct health information during emergencies). A possible explanation for less data protection by disabled and unwell people may be that perceived benefits are higher for these groups of people compared to well people. However, the question remains whether benefits (especially societal benefits) of personal data disclosure with healthcare professionals is visible to healthy people. Therefore it is expected that perceived benefits will be higher for people with CVD complaints which will result in more willingness to disclose personal information.
Furthermore, Miltgen & Smith (2015) add that consumers are less cautious in their privacy behavior when they trust in the privacy legislation of the nation. Results of the ten semi-structured interviews among ten University of Indiana students in the U.S. by Udoh & Alkharashi (2017) showed similar results. While American students appeared to be concerned with privacy, a particular student from Greece indicated to not have concerns. The respondent stated that in his origin country, cyber-attacks rarely occur. However, the student did indicate that privacy policies in the U.S. must be made clearer. Although there are strict privacy restrictions when it comes to privacy within healthcare in the Netherlands (KNMG, 2018), it is not clear whether non-patients are also aware of the privacy protection regulations. One can assume that patients with a heart disease are more likely to be involved in research and hospital protocols and therefore have more trust in the privacy legislation of the nation compared to people who have not been involved in medical research. This would be in line
12 with results of Beranek Lafky and Horan (2011) who found that patients perceive less concerns and are less secure when it comes to information disclosure. It is expected that individuals with CVD complaints perceive more coping capabilities and benefits compared to individuals without CVD complaints and consequently are more likely to disclose personal information. This leads to the formulation of the first hypotheses of this study:
H1
The positive relationship between CVD complaints and information disclosure is mediated by (1) perceived benefits and (2) perceived coping capability.
H1a: An increase (decrease) in CVD complaints will increase (decrease) perceived benefits which positively (negatively) mediates the effect on information disclosure.
H1b: An increase (decrease) in CVD complaints will increase (decrease) perception of coping capability which positively (negatively) mediates the effect on information disclosure.
Smartwatch brand trust
Brand trust has been defined as “the extent to which the consumer believes that the brand accomplishes its value promise, and brand intentions, which is based on the extent to which the consumer believes that the brand would hold consumers’ interests ahead of its self-interest when unexpected problems with the consumption of the product arise” (Delgado-Ballester, 2004, p. 575). Also, brand trust is being established over time (Rempel, Holmes, & Zanna, 1985). It takes individuals several experiences and other interactions before building such trust. Results from research of Wottrich, Verlegh, & Smit (2017) have shown that people with high brand trust are more likely to disclose personal information. This finding aligns with the study of Mesch (2012) who found that the willingness to share information is being
13 influenced by someone’s trust in a brand. Additionally, Waiguny, Nelson and Marko (2013) found that familiar brands are less likely to be vulnerable to the influence of violent content in games. This may also provide an explanation for the fact that high brand trust results in individual’s likeliness to disclose information. Smartwatch brand trust and its effect on individuals’ willingness to disclose information is considered relevant to this study since cardiologists among the healthcare professionals are debating on what devices to use to measure health data of which smartwatches are a relevant option (appendix 1). Udoh & Alkharashi (2017) conducted semi-structured interviews among ten Indiana University students in the U.S. showing that respondents preferred Apple smartwatches over other brands such as Samsung, HP and FitBit due to higher privacy expectations. This finding is interesting to take into consideration although the author’s sample size is relatively small (N=10). Also, according to privacy awareness literature, privacy awareness also include one’s trust in technology. Considering findings of Waiguny et al. (2013) and Rempel et al. (1985) one may assume that brand trust will be higher for familiar brands compared to unfamiliar brands since trust establishes over time. Apple will be considered as a familiar brand while Zeblaze (i.e. a relatively new smartwatch brand established in China) will be considered as an unfamiliar brand in this study, since Apple has been on the market for over twenty years and Zeblaze for four years. The following hypotheses have been formulated to test whether brand indeed affects one’s perceived benefits and perceived coping capabilities.
H2
The positive relationship between smartwatch brand trust – which will be higher for Apple than for Zeblaze - and information disclosure is mediated by (1) perceived benefits and (2) perceived coping capability.
14 H2a: An increase (decrease) in smartwatch brand trust will increase (decrease) perceived benefits which positively (negatively) mediates the effect on information disclosure.
H2b: An increase (decrease) in smartwatch brand trust will increase (decrease) perception of coping capability which positively (negatively) mediates the effect on information disclosure.
Reputation of information receiver
Ellen Janssens (appendix 1) assumes that privacy behavior is influenced by the party to which consumers share data with. She explained that patients tend to share personal data with cardiologists, but not necessarily with health insurance companies. Research by Xie et al. (2006) stated that Doney and Cannon (1997) defined reputation in the marketing literature as: “the extent to which firms and people in the industry believe a firm is honest and concerned about its customers.” (p.63). Furthermore, Laveist, Isaac, & Williams (2009) found that mistrust of patients in healthcare resulted to be a critical predictor in the health technology adaption. These patients tend to lack health care services as a result of a low health privacy capital. Past researches (Acquisti, Brandimarte, & Loewenstein, 2015; Bansal, Zahedi, & Gefen 2016) focused on the relationship between privacy behavior and context. Results showed that people were more likely to disclose personal and sensitive information to websites with less formal contexts rather than to very professional (and trustable) looking websites. This finding is interesting, although it is not directly applicable to the setting of this research since the parties are all professional and only vary in purposes. Nevertheless, varying reputations (from non-profit to commerce) may influence one’s perceived benefits and perceived coping capabilities. Therefore two different data receivers have been added to this study to compare differences of reputation. Erasmus Medical Center will be compared to Menzis (health insurance company).This leads to the third hypotheses of this study:
15 H3
The positive relationship between positive reputation of data receiver – which will be higher for ErasmusMC than for Menzis - on information disclosure is mediated by (1) perceived benefits and (2) perceived coping capability.
H3a: An increase (decrease) in the positive reputation of the data receiver will increase (decrease) perceived benefits which will positively (negatively) mediate the effect on information disclosure.
H3b: An increase (decrease) in the positive reputation of the data receiver to will increase (decrease) an individual’s perception of coping capability which will positively (negatively) mediate the effect on information disclosure.
3. Research design
The formulated hypotheses are presented in the conceptual framework (figure 3.1). This study will continue referring to: CVD complaints as ‘CVD’, smartwatch brand trust as ‘SBT’, reputation of data receiver as ‘RDR’, perceived benefits as ‘PB’, perceived coping capability as ‘PCC’ and information disclosure as ‘ID’.
Figure 3.1 Conceptual framework
The hypotheses are formulated on a general level. Meaning they are formulated to test whether CVD, SBT and RDR affect PB, PCC and eventually ID. However, this study also
16 included two smartwatch brands (i.e. Apple and Zeblaze) and two data receivers (i.e. Erasmus Medical Center and Menzis health insurance) to test whether means vary between groups. Hence, to test whether PB, PCC & ID differ between CVD, SBT and RDR between groups.
3.1 Sample and response
Dutch citizens are the population of this study, specifically: citizens without CVDs and citizens suffering from CVDs. These subgroups are necessary to measure the expected different outcomes on willingness to disclose information. It was not possible to use stratification sampling during this study, due to the decision of the Dutch Heart Foundation’s patients’ organization (Harteraad) to deny accessibility to CVD patients. The Harteraad found that this study was not directly relevant and applicable to their audience, since respondents were exposed to scenarios in which they were asked to answer questions related to Menzis and Erasmus Medical Center. They indicated that their audience may not have a relation with these organizations. The Dutch Heart Foundation was willing to send an invitation to their connections through a newsletter and social media (e.g. Facebook and Twitter). However, this would have taken many months to organize due to the Heart Foundation’s content planning schedule and therefore would exceed the time schedule of this study. These constraints have resulted in the decision to use quota, volunteer and convenience sampling.
Patients suffering from CVDs were reached through invitations by the AFIP foundation (afiponline.org) and a patients’ community on Facebook (LetTheBeatGoOn). Also, an article calling out patients to participate was posted on the AFIP website. In addition to this article, employees of ErasmusMC reached out to patients with an invitation to fill out the questionnaire who came to see a cardiologist. In total 126 respondents suffering from CVDs completed the survey. People without CVD complaints were collected through convenience or availability sampling. These respondents were reached through social media
17 (i.e. Facebook and Instagram), Email and direct personal approaches. In total 127 respondents without CVD complaints completed the survey.
Additionally, this study aimed to collect 400 (completed) responses. Specifically, 200 healthy citizens and 200 citizens suffering from CVDs. Due to the Harteraad denying access, the total of completed responses is 253. Previously Park & Chung (2017) conducted similar research and sampled within an American Cancer community and received a response rate of 26%. A similar response rate for the invitations that had been sent out to patients through the AFIP foundation was expected. The foundation has 103 patients in their mailing list and on average the response rate was 38.1%. Furthermore, 330 respondents participated in the survey of which 77 (23%) did not complete the survey.
3.2 Survey design and measures
Qualtrics was used to design the questionnaire and collect responses. In appendix 5 an overview of the survey is provided. Firstly, the questionnaire asked respondents whether they suffer from CVD complaints or not. This was measured through asking whether the respondent has or had CVD complaints, resulting in a nominal outcome (either ‘yes’ or ‘no’). Then randomization was used in the survey to expose respondents randomly to scenarios in which they had to answer questions related to one of the four situations: (1) ErasmusMC as data receiver and Apple as smartwatch brand, (2) ErasmusMC as data receiver and Zeblaze as smartwatch brand, (3) Menzis as data receiver and Apple as smartwatch and (4) Menzis as data receiver and Zeblaze as smartwatch brand. Table 3.1 presents an overview of the scenarios a respondent had a (random) chance to be exposed to.
Table 3.1 Survey Scenarios
Scenario Combination
1 Erasmus and Apple
2 Menzis and Apple
3 Erasmus and Zeblaze
18 Two instances were used to create different situations in both SBT and RDR groups. Respondents were exposed to a description including a call-out to participate in research. ErasmusMC was described as an organization aiming to improve society’s (heart) health (i.e. societal benefit). In contrast, Menzis was presented as an organization aiming to gain more insights in society’s (heart) health in order to successfully adapt their services to these findings. Both descriptions mentioned that respondents were able to track and access their own data (i.e. personal benefit). Therefore, ErasmusMC contained over two benefits and Menzis over one. Next, respondents received a description in which they were introduced to either one of the smartwatches, which ErasmusMC or Menzis wanted to use to conduct the research with. Apple was described as a well-known brand in contrast to Zeblaze, which was described as an unknown upcoming brand from China. The descriptions included information on who would have ownership of the respondent’s data in order to ensure privacy awareness among groups were equal. Not solely ErasmusMC and Menzis would have ownership over the respondent’s data, but also the smartwatch brands.
Within these scenarios, respondents were asked to indicate whether they agreed with statements or not which were used on a 7-items Likert scale to measure SBT and RDR. During this study a 7-items Likert scale was preferred over a 5-items Likert Scale since research by Lewis (1993) has shown stronger correlations with observed significance for 7-point scales compared to 5-7-point scales. Additionally, no more than seven items were chosen, since we cannot encompass more than seven items in our memory (Colman, Norris, & Preston, 1997). Therefore, all used scales are extended to 7 points. Appendix 3 provides an overview of all measures used during this study.
19 A pre-test was conducted before sending out the questionnaire to respondents. Eight respondents got to review the questions and add comments where necessary. Eventually a view spelling mistakes were corrected and some questions were formulated more clearly.
SBT was measured through the three itemed scale of Kuang-Wen Wua, Shaio Yan Huang, David C. Yen (2012) (e.g. ‘I trust the smartwatch that protects personal information’). The 7-points on the Likert-Scale ranged from 1= ‘strongly disagree’ to 7= ‘strongly agree’. The Cronbach’s alpha that was found for this scale is .74.
RDR was measured through the four itemed scale by Ponzi, L. J., Fombrun, C. J., & Gardberg (2012) which measures reputation of the data receiver (e.g. ‘[Company] is a company that I admire and respect’). The 7-points on the Likert-Scale ranged from 1=‘strongly disagree’ to 7= ‘strongly agree’. The Cronbach’s alpha for this scale is .95.
PB was measured through the 7-point Likert scale of Inman, J., & Nikolova (2017) with a Cronbach’s alpha of .73 (e.g. ‘The value I gain from using this technology is worth the information I give away’). The scale consists out of three items, including statements regarding privacy risks. Privacy risks were included into the scale since it grounds the perception of the weight of benefits. The points on the scale range from 1= ‘strongly disagree’ to 7= ‘strongly agree’.
PCC – also known as confidence in the literature - was measured using the Likert-Scale by Park & Chung (2017) (e.g. ‘How confident are you that you have some say in who is allowed to collect, use, and share your medical information?’). The researchers used this scale with only one statement. However, the statement covers three separate elements. Therefore these elements have been split up into three statements for this study. The Cronbach’s alpha is .74 and the 7-points on the scale ranged from 1= ‘not very confident’ to 7= ‘very confident’.
20 The 5-point Likert scale of Kuang-Wen Wua, Shaio Yan Huang, David C. Yen, (2012) was used as input to measure willingness to disclose information (e.g. ‘You are willing to provide personal information to the website’). Also for this scale the 5-point rating scale was replaced for a 7-point rating. The Cronbach’s alpha is .823 and the 7-points ranged from 1= ‘strongly disagree’ to 7= ‘strongly agree’. Although the original scale consisted out of two items, one item (‘You are forced to provide personal information to the websites’) was removed since it was not related to ID in the context of this study.
4. Results
First of all, the statistical procedure of this study is discussed along with the preparation of data for analyses. Descriptive statistics are discussed along with the reliability of scales and new variables that were created. After, the results of the hypotheses tests are presented.
4.1 Statistical procedure
First of all, data has been prepared for analyses. Descriptive frequencies was run in order to find missing values and errors (table 4.1). Although no errors were found, there were 77 missing cases which make out 23% of the data. These cases have been excluded list wise from the dataset which resulted in 253 useable cases for analyses. Variables which were measured for all respondents (i.e. ‘CVD’, ‘PB’, ‘PCC’ and ‘ID’) have been included in the descriptive analysis, since the survey included a randomizer to measure four different scenarios of ‘SBT and ‘RDR’. Additionally, for almost all variables the skewness and kurtosis are acceptable since the scores vary between -2 and +2 (George & Ph, n.d.). Hence, normality of the distributions is assumed. Nevertheless, considering the individual scenario groups, the reputation of ErasmusMC holds a high Kurtosis value (≥2). In other words, the distribution of the reputation of Erasmus peaks at a certain mean value which makes the distribution which makes this population non-normally distributed. This may be problematic while interpreting
21
Table 4.1 Descriptive Statistics
N Mean Std. Deviation Skewness Kurtosis
CVD 253 1.502 .50099 -.008 .153 -2.016 .305 PB 253 4.177 .66831 -.626 .153 .504 .305 PCC 253 3.474 1.56287 .173 .153 -.955 .305 ID 253 1.482 .50067 .072 .153 -2.011 .305 RDR 253 4.717 1.4178 -.602 .150 -.118 .299 SBT 253 3.714 1.60827 .004 .150 -1.074 .300 RepErasmus 128 5.5352 .97848 -1.313 .214 3.384 .425 RepMenzis 125 3.91 1.30539 -.158 .217 -.125 .43 TrustApple 106 4.4403 1.37671 -.542 .235 -.641 .465 TrustZeblaze 147 3.2245 1.52412 .436 .2 -.66 .397 results for mean differences between ErasmusMC and Menzis, since the distributions are not similar. However, combing the results with the results from the reputation of Menzis, resulted in a normal distribution of the population. This population was used for regression analyses to test whether RDR affects ID.
Recoding
Three variables have been recoded. One scale item of PB was formulated in a negative way to increase the likelihood of respondents filling out honest answers. The scale of ‘I think the
risks of my information disclosure will be greater than the benefits gained from the use of this technology’ has been recoded in a counter-indicative manner (i.e. 1= ‘strongly agree’ to 7=
‘strongly disagree’).
Furthermore, ‘yes’=1 and ‘no’=2 on whether the respondent has CVD complaints, has been switched to ‘yes’=2 and ‘no’=1 in order to improve the interpretation of correlations. The latter option will result in positive correlation and coefficient outcomes instead of negative ones. This variable was then labeled as ‘CVD’. Finally, ID was measured through a scale ranging from 1= ‘strongly disagree’ to 7= ‘strongly agree’. These results have been divided into binary categories (scale rating ≤ 3 =’no’, scale rating >3= ‘yes’) resulting in a dichotomous outcome since the scale of ID solely consisted out of one item. Therefore, means
22 of the scale could not be computed which made the outcome unsuitable for a linear regression analysis. Recoding the ID into ‘yes’ and ‘no’ makes the outcome appropriate for a logistic binary regression model.
New variables were created in addition to the recoding process. These new variables were necessary in order to divide the different scenarios (groups) that have been presented to respondents in the survey. This was done for SBT and RDR groups, but also for these two groups combined to test the total effect of scenario combinations on ID. The RDR variable holds the reputation for Erasmus (=1) and for Menzis (=2). SBT includes smartwatch brand trust for Apple (=1) and for Zeblaze (=2). The RDR and PBT variables were also hold separate in order to perform mean analyses between the two instances within these groups, which otherwise could not have been done. Next the variable ‘Scenario’ was created and represents a combination of SBT and RDR to which the respondent was exposed. Respondents are therefore divided into a scenario of either Erasmus and Apple (=1), Menzis and Apple (=2), Erasmus and Zeblaze (=3) or Menzis and Zeblaze (=4). These new variables enabled the possibility to analyze differences in means between scenarios through a one-way ANOVA test which was needed to test mean differences in ID.
Reliability
A reliability test was run for the scale measures of SBT (three items), RDR (four items), PB (three items) and PCC (four items) after checking data for errors and recoding variables. All scales for SBT and RDR have been tested on reliability, since these variables include different scenarios. Respondents were exposed to one out of four scenarios and therefore four scales of both SBT and RDR have been included in the reliability analysis as can be read from table 4.2. All Cronbach’s alpha scores were sufficient (i.e. α >.70). In all cases the scores exceeded α >.90 except for one scale which measured SBT for the situation where respondents were asked questions about ErasmusMC and Apple. Also, this was the only scale where the
23 Cronbach’s’ alpha score would improve by α >.10 in case of removal of one scale item. However, all scale items were kept even though the Cronbach’s’ alpha would improve after deleting that particular item, since the overall score exceeded α >.80. Furthermore, all corrected-item totals hold a sufficient score of the scale (i.e. above α>.30). The results indicate that all scales were able to produce consistent results and therefore yield consistent findings.
Table 4.2 Reliability of scale items
Variable α RDRErasmusApple .93 RDRErasmusZeblaze .95 RDRMenzisApple .93 RDRMenzisZeblaze .95 SBTAppleErasmus .88 SBTAppleMenzis .91 SBTZeblazeErasmus .94 SBTZeblazeMenzis .93 PB .90 PCC .96 Computing variables
Mean totals were computed for all variables which are presented in table 4.3 in order to make them suitable for regression analyses and one-way ANOVA analyses. This resulted in total means for each component of RDR (i.e. reputation for ErasmusMC and Menzis within four scenarios) and SBT (i.e. smartwatch brand trust for Apple and Zeblaze within four scenarios). In total four different variables holding mean totals for RDR were computed and four different variables holding mean totals for SBT were computed. The hypotheses of this study intend to test general effects of SBT and RDR (i.e. RDR scenarios combined and SBT scenarios combined) on ID and therefore it was necessary to compute mean totals for each individual situation into a new variable, including total means for SBT and RDR. Mean totals for reputation of ErasmusMC have been merged with the mean totals of the reputation of Menzis in order to test the general effect of RDR on ID. This new variable was then labeled as
24 ‘RDR’. The same was done for mean totals regarding SBT of Apple and Zeblaze. This new variable was labeled as ‘SBT’.
Correlations
Table 4.3 provides an overview of correlations between this study’s variables. Some significant correlations are found of which most have a significance level of <0.01 (2-tailed). Overall, most correlations have a (positive) moderate effect. Meaning that an increase in a specific variable (e.g. CVD) will also result in an increase of another specific variable (e.g. ID). However, it must be noted that these correlations and coefficients do not say anything about causal relationships between variables.
Furthermore, the results indicate that most independent variables of this study are correlated, which may cause multicollinearity. This is barely the case for CVD which seems to be not correlated to RDR (β=.089 p > .05) and very weakly correlated to SBT (β=.128 p < .05). However, RDR and SBT seem to be positively correlated (β=.341 p < .01) as well as PCC and PB (β=.447 p < .01).
Although the correlations are not problematically strong (i.e. β=.>.50), they do indicate that two independent variables may have similar measures and should therefore possibly be included into one measure (i.e. one independent variable). Multicollinearity may cause a model to change drastically once removing a correlated variable from the analysis and can make it challenging to assess the importance of an independent variable explaining variation (Mark Saunders, Philip Lewis, 2016). Nevertheless, since correlations are not exceeding .50 this is not assumed for this study although caution must still be taken while interpreting results.
25
4.2 Analyses outcomes
As mentioned before, the hypotheses of this study are formulated on a general level (i.e. the general effect of CVD, RDR and SBT on ID) and on ‘scenario’ level (i.e. whether there are significant differences in means between groups of CVD, RDR and SBT). The hypotheses are tested through (1) logistic (binary) regression analyses to test effects of independent and mediating variables on ID and (2) linear regression analyses to test effects of independent variables on the mediating variables. Furthermore, one-way ANOVA tests were performed to test mean differences between different scenarios for SBT and RDR. Also, Mann-Whitney U tests were executed to test whether different means were found in ID for CVD, SBT and RDR.
Furthermore, mediation was tested in multiple steps since the dependent variable (i.e. ID) was made binary which made mediation testing through PROCESS by Andrew F. Hayes (2015) inappropriate. Therefore, the mediating effect of PB and PCC were tested by (1) testing the effect of the independent variables (i.e. CVD, SBT and RDR) on expected mediators (i.e. PB and PCC), (2) testing the effect of independent variables on ID individually and in while controlling for the expected mediators and finally (3) testing for significance through the Sobel Test (Sobel, 1982).
Table 4.3 Means, Standard Deviations and Correlations
Variables M SD 1 2 3 4 5 6 1. CVD 1.502 .50099 2. ID 1.4822 .50067 .423** 3. PB 4.1779 .66831 .242** .403** 4. PCC 3.4743 1.56287 .266** .354** .447** 5. RDR 4.7322 1.40848 .089 .374** .426** .405** 6. SBT 3.7339 1.58012 .128* .342** .490** .461** .341** ** Correlation is significant at the 0.01 level (2-tailed).
26 Overall the mediation formula is: c= c’+ab (Kenny, 2018). In this formula c stands for the total effect of the independent variable on the dependent variable including mediation, c’ refers to the direct effect of the independent variable on the dependent variable while being controlled for the mediator, a refers to the direct effect of the independent variable on the expected mediator and b stand for the direct effect of the mediating variable on the dependent variable (while including the independent variable as covariate into the model). Generally coefficients from multiple regression analyses are suited to be included into the formula in order to test for mediation. However, in this particular study, coefficient scales of independent (X), mediating (M) and dependent variables (Y) differed from one another. Coefficients derived from linear regression analyses were not appropriate to directly compare to coefficients derived from logistic regression analyses. Therefore, additional calculations were made to enable calculations for total mediation effects. Calculation tools (i.e. SPSS syntax and a calculation spreadsheet) provided by Herr (2016) were used. The next section discusses statistics which were necessary to calculate and convert outcomes.
Converting coefficients
First of all, standard deviations of the independent, mediating and dependent variables were derived from table 4.1. These standard deviations were necessary to make coefficients comparable across equations. The formula used is: comparable a = a * 𝑆𝑆𝑆𝑆𝑋𝑋/𝑆𝑆𝑆𝑆𝑀𝑀′, comparable b= b *𝑆𝑆𝑆𝑆𝑀𝑀/𝑆𝑆𝑆𝑆𝑌𝑌", comparable c= c * 𝑆𝑆𝑆𝑆𝑋𝑋/𝑆𝑆𝑆𝑆𝑌𝑌′, comparable c’= c’ * 𝑆𝑆𝑆𝑆𝑋𝑋/𝑆𝑆𝑆𝑆𝑌𝑌′. The primes are added to the notations since, for instance, M can either be a predictor or a dependent variable. A prime stands for the variable being dependent. The double prime means that that the dependent variable has two predictors. Next the variances between standard deviations were solved through the following equations: Var(Y') = c2 * Var(X) + 𝜋𝜋²/3, Var(M') = a2 * Var(X) + 𝜋𝜋²/3, Var(Y") = c'2 * Var(X) + b2 * Var(M) + 2*b*c'*Cov(X,M) + 𝜋𝜋²/3 (Kenny, 2018). Furthermore, Kenny formulated equations for comparable standard errors which are:
27 SE(comp a) = SE(a) * SD(X)/SD(M'), SE(comp b) = SE(b) * SD(M)/SD(Y"), SE(comp c) = SE(c) * SD(X)/SD(Y'), SE(comp c') = SE(c') * SD(X)/SD(Y").
Next, the correlation coefficients between X and M were retrieved from table 4.3 and inserted into the equation to calculate variances between standard deviations. Also, regression coefficients and standard errors for effects of (1) X on M, (2) M on X controlling for the independent variable, (3) X on Y without controlling for M and (4) X on Y while controlling for M were used in the equation to make path a, b, c and c’ comparable. These coefficients are presented in table 4.4, which sums up all coefficients for the hierarchical linear regression analysis that was performed to get effects of X on M, and in table 4.5, which provides the coefficients that were found after executing the logistic regression analysis that was necessary to analyze the effect of X and M on Y. As has been mentioned before, Herr (2016) added the necessary equations into a spreadsheet so that the calculations could be done automatically instead of manually. Besides providing equations to make coefficient pathways comparable, Herr’s spreadsheet also calculates the Sobel score and Sobel Z score to test whether mediation was found significant.
After running linear and logistic regression models (table 4.4 and 4.5), results showed that CVD, RDR and SBT affected ID directly and not solely indirectly. These independent variables formed the scenarios of this study which may affect outcomes in PB, PCC and ID. Therefore, these variables were included as covariates in both regression models to receive accurate coefficients for the relation between the variables on ID.
Also, it is notable that PCC does not affect ID significantly within the model (p>.10) Therefore H1b, H2b, H3b are rejected and only mediation for PB was further tested in the next sections.
28
Mediation effect of PB between CVD and ID
First of all, the effect of CVD on PB was tested. RDR and SBT were added to the (hierarchical) linear regression model as covariates. Both these variables represent the scenario to which respondents were exposed and therefore may affect variation in PB. The results in table 4.4 show that in model 1 CVD explains 5.8% in the variance of PB. This is a relatively small percentage. The R² increased to .340 after adding SBT and RDR into the model. In other words, the predictive power of the model increased by .283. Overall a weak
Table 4.4 Linear Regression CVD, SBT and RDR on PB
R R² R² change B SE β t Model 1 .240 .058*** CVD .320 .081 .240*** 3.940 Model 2 .583 .340*** .283*** CVD .215 .069 .161** 3.117 SBT .156 .023 .371*** 6.727 RDR .131 .026 .279*** 5.090
Note: statistical significance: **p <.01; ***p <.001
Table 4.5 Logistic regression models predicting ID
β SE Exp(B) p Model 1* CVD 2.000 .323 7.386 .000 SBT .351 .105 1.421 .001 RDR .643 .140 1.903 .000 Model 2** CVD 1.863 .331 6.444 .000 SBT .235 .117 1.265 .044 RDR .568 .146 1.764 .000 PB .619 .303 1.857 .041 PCC .077 .119 1.080 .515 * X²=96.300, df=3, p <.001 **X²=101.903, df=5, p <.001 Nagelkerke R² 41.9% 43.9%
Hosmer & Lemeshow test p=.880 p=.695
29 positive coefficient was found between CVD and PB (β=.161, p<0.01). In other words, when someone has CVD complaints perceived benefits will increase by .161.
Next, a logistic analysis was run to test CVD against ID. Also in this case, SBT and RDR were added as control variables to the model. Model 1 showed a positive coefficient between CVD and ID (β= 2.000 p<0.001) (table 4.5). Also, Model 1 was found to be significant and has a goodness of fit of 41.9% and was able to predict 72.9% of the cases. Adding PB and PCC to the model the goodness of fit increased to 43.9% and the prediction accuracy increased to 75.7%. This indicates that the addition of the expected mediators did not add a notable amount (i.e. only 2%) to the goodness of fit. In this model it appears that for a unit increase in CVD it is expected that the log odds of ID will increase by 2.000. This correlation coefficient decreased in model 2, after inserting PB and PCC into the model (β=1.863, p <.001). This indicates PB partly and weakly mediates the effect of CVD on ID, since the effect remained significant. The standard deviations ( 𝑆𝑆𝑆𝑆𝑋𝑋= .50049, 𝑆𝑆𝑆𝑆𝑀𝑀= .66861, 𝑆𝑆𝑆𝑆𝑌𝑌= .49064), correlation coefficients ( 𝛽𝛽𝑋𝑋,𝑀𝑀 = .24) and standard errors ( 𝛽𝛽𝑎𝑎=.215, 𝑆𝑆𝑆𝑆𝑎𝑎= .069, 𝛽𝛽𝑏𝑏= .619, 𝑆𝑆𝑆𝑆𝑏𝑏= .303, 𝛽𝛽𝑐𝑐=2 𝑆𝑆𝑆𝑆𝑐𝑐= .323, 𝛽𝛽𝑐𝑐′=1.863, 𝑆𝑆𝑆𝑆𝑐𝑐′=.331) were all added
into Herr’s spreadsheet which provided a total effect of CVD on ID of .43 (table 4.6). Considering the fact that this effect is almost equal to the direct effect of CVD alone, there seems to be a mediation on first sight. Although this coefficient indicates a positive moderate effect, Baron and Kenny’s (2018) proportion of M’s (i.e. PB’s) effect was found to be only 2.2%. Hence, only 2.2% of CVD’s effect on ID is due to mediation. Finally the Sobel test was run and found an insignificant Z value (table 4.6). Therefore, there is no evidence CVD’s effect on ID is being mediated by PB. Hence, CVD does affect ID, yet not through mediation of PB.
30 Again, considering the results from table 4.5, PCC does not seem to significantly affect ID while controlling for PB, RDR, SBT and CVD. Therefore the steps needed to test mediation were not performed for CVD’s effect on ID through PCC.
Next a Mann-Whitney U test was performed to test for significant mean differences between respondents with and without CVD complaints. For this test the original (ordinal) ID scale was used since testing the (binary) new variable for ID for mean differences was limited to tests, such as the chi² test which are less convenient to test means between groups with. Table 4.7 shows that the mean difference between individuals with and without CVD complaints is 66.4 (160.7-94.3, p<.001). Hence, individuals with CVD complaints are more willing to disclose information.
Table 4.7 Mann-Whitney U results for mean differences in ID by CVD groups
N Mean Sum of Ranks
No CVD 127 94.3 11976
CVD 126 160.7 20409
Total 253
Mann-Whitney U 3848
Asymp. Sig. (2-tailed) .000
Table 4.6 Results mediation of PB between CVD and ID
β SE Comparable a .059221 .019006 Comparable b .187269 .091668 Comparable c' .421901 .074959 Comparable c .483167 .078032 Outcome ab+c'= .432991
Percentage of Effect M (Kenny, 2018) .022953
Sobel .006491
Sobel Z 1.708452
31 Results show that CVD positively and moderately affects ID. Considering the comparable coefficients pathways, having CVD complaints will result in an increase in ID of .48. When an individual’s CVD complaints increase by 1 unit willingness to disclose information will increase by .483167. However, there has been no evidence found that this relationship is mediated by PB or by PCC. This means that both hypothesis 1a and 2b are not supported, yet hypothesis 1 is partly supported. Also, the Mann-Whitney U test found that individuals with CVD complaints are significantly more willing to disclose information.
Mediation effect of PB between SBT and ID
The same steps to calculate mediation were used to test mediation of PB and PCC between SBT and ID. First a (hierarchical) linear regression analysis was executed to measure the effect of X (i.e. SBT) on M (i.e. PB). Also in this analysis SBT was controlled for CVD and RDR, since both other independent variables may affect PB. The output, which is presented in table 4.4, indicates that when SBT increases by 1 PB will increase by .371 (β=.371, p <.001). This effect is relatively weak and positive. As has been mentioned before, the model explains 34% of the variance in ID which means that 66% of the variance is explained by other factors. Next, logistic regression analyses were run in order to test the direct effect of SBT on ID and to test the effect of PB on ID while controlling for PCC, RDR and CVD. A positive (weak) relationship was found between SBT and ID (β= .351, p <.05). After entering PB and PCC into the model, this coefficient decreased by .116 (β=.235, p <.05). In other words, a unit increase in smartwatch brand trust will expect to result in a log odds increase in information disclosure of .351. This decrease indicates a partly mediation caused by either PB, PCC or both, since SBT’s effect on ID remained significant. Since PCC does not affect ID significantly, mediation will only be tested for PB. The standard deviations (𝑆𝑆𝑆𝑆𝑋𝑋= 1.60827, 𝑆𝑆𝑆𝑆𝑀𝑀=.66861, 𝑆𝑆𝑆𝑆𝑌𝑌= .49064 ), correlation coefficients ( 𝛽𝛽𝑋𝑋,𝑀𝑀= .492 ) and standard errors ( 𝛽𝛽𝑎𝑎= .371, 𝑆𝑆𝑆𝑆𝑎𝑎= .156, 𝛽𝛽𝑏𝑏= .619, 𝑆𝑆𝑆𝑆𝑏𝑏= .303, 𝛽𝛽𝑐𝑐=.351, 𝑆𝑆𝑆𝑆𝑐𝑐= .105, 𝛽𝛽𝑐𝑐′=.235,
32 𝑆𝑆𝑆𝑆𝑐𝑐′=.117) were all added into Herr’s spreadsheet which provided a total effect of SBT on ID
of .26 (table 4.8). This is a relatively weak total effect and is indeed almost equal to the direct effect of SBT on ID. The general Sobel score was found significant (p<0.05), however, the Sobel Z score is not (p<1.95). Therefore, no evidence was found of PB’s mediating role between the relationship of SBT and ID.
Table 4.8 Results mediation of PB between SBT and ID
β SE Comparable a .312481 .131394 Comparable b .213799 .104655 Comparable c' .19524 .097205 Comparable c .297162 .088895 Outcome ab+c'= .262049
Percentage of Effect M (Kenny, 2018) .0224821
Sobel .0043112
Sobel Z 1.549657
Note: Sobel is significant at <.05, Sobel Z is significant at >1.95
Finally, differences in means for the smartwatch brand groups (i.e. Apple and Zeblaze) were tested through a one-way ANOVA, since the computed means of SBT made an ANOVA test appropriate to use. The new variable groups were used (1= ‘Apple’ and 2= ‘Zeblaze’) to measure differences in SBT. The results are presented in table 4.9 and indicate that there is significant difference found in means between Apple and Zeblaze (𝑀𝑀𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴= 4.4969, 𝑀𝑀𝑍𝑍𝐴𝐴𝑏𝑏𝐴𝐴𝑎𝑎𝑍𝑍𝐴𝐴=3.1645, p<.001). Additionally, the Levene statistic indicates that homogeneity is
found within the distribution of the groups (p>.05) which contributes to the reliability of the ANOVA outcomes. This result means that on average individual’s trust Apple more than Zeblaze. Considering the logistic regression analysis SBT has a direct effect on ID and therefore it’s expected that individuals who evaluated Erasmus are more willing to disclose information.
33
Table 4.9 One-way ANOVA test for SBT
Descriptives N M SD SBT Apple 108 4.4969 1.37361 Zeblaze 145 3.1645 1.53499 Total 253 3.7137 1.60827 ANOVA SS DF MS F Sig. SBT Between Groups 112.699 1 112.699 52.102 .000 Within Groups 562.387 251 2.163 Total 675.086 252 Homogeneity of variances
Levene Statistic df1 df2 Sig.
SBT .917 1 253 .339
To test whether this expectation is true, another Mann-Whitney U test was run in order to test for significant mean differences in ID for the smartwatch brands. The results for this test are shown in table 4.10.
Table 4.10 Mann-Whitney U results for mean differences in ID by SBT groups
N Mean Sum of Ranks
Apple 108 142.48 14960.5
Zeblaze 145 116.94 17424.5
Total 253
Mann-Whitney U 6249.500
Asymp. Sig. (2-tailed) .006
The output of the Mann-Whitney U test indicates that there is indeed a difference in willingness to disclose information for respondents that were either exposed to a scenario related to Apple or to Zeblaze. Scenarios which included Apple as smartwatch brand hold a higher mean for ID than for scenarios containing Zeblaze (𝑀𝑀𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴= 142.48, 𝑀𝑀𝑍𝑍𝐴𝐴𝑏𝑏𝐴𝐴𝑎𝑎𝑍𝑍𝐴𝐴=116.94, p<.005). Combining these insights with the fact that SBT has a (relatively weak) direct positive effect on ID, one can state that an increase in a unit of SBT will increase ID by .297162. However, this effect is neither mediated by PB nor by PCC. Therefore both hypotheses 2a and 2b are rejected, yet hypothesis 2 is partly supported.
34
Mediation effect of PB between RDR and ID
The linear regression that was run shows the effect of RDR on PB while controlling for SBT and CVD. A positive, yet relatively weak relation, was found (β=.279, p<.001). PB will therefore increase by .279 for every time RDR increases by 1. Next the effect of RDR on ID and the effect of PB on ID was analyzed. These results are presented in table 4.4 as well. The predictor RDR has a positive effect on ID (β=.643, p<.001). After entering PB and PCC into the model this effect decreased by .075 (β=.568, p<.001). Therefore, it is expected that a unit increase in the reputation of the data receiver will result in a log odds increase in information disclosure of .568. These findings indicate that PB has a partly mediating role between RDR and ID since RDR’s effect on ID remained significant. The standard deviations (𝑆𝑆𝑆𝑆𝑋𝑋= 1.41789, 𝑆𝑆𝑆𝑆𝑀𝑀= .66861, 𝑆𝑆𝑆𝑆𝑌𝑌= .49064 ), correlation coefficients (𝛽𝛽𝑋𝑋,𝑀𝑀 = .427) and standard errors ( 𝛽𝛽𝑎𝑎=.213088, 𝑆𝑆𝑆𝑆𝑎𝑎= .019858, 𝛽𝛽𝑏𝑏= .197077, 𝑆𝑆𝑆𝑆𝑏𝑏= .096469, 𝛽𝛽𝑐𝑐= .449098 𝑆𝑆𝑆𝑆𝑐𝑐= .097782, 𝛽𝛽𝑐𝑐′=.383499, 𝑆𝑆𝑆𝑆𝑐𝑐′=.098575) were all added into Herr’s spreadsheet which provided a total effect of RDR on ID of .43 (table 4.11). This positive and moderate total effect is almost equal to the direct effect of RDR on ID (i.e. β=.449098). In total, the mediation has a proportion of 9.3% within this total effect. Furthermore, the scores derived from the significance indicate that the mediation by PB is significant (Sobel Z >1.95). Therefore H3a is partly supported. Once RDR increases by one, ID will increase as well by .43 through a direct effect but also a partly mediation by PB.
