DCEs can be used to quantify preferences for products or services and are increasingly used within the health care setting [16,17]. This method is based on the Random Utility Theory (RUT) and requires respondents to answer several choice tasks. Such tasks present two or more profiles of a specific good or service. The profiles are described based on their characteristics (ie, attributes) with varying levels [18-20]. Within each choice task, respondents are asked to choose the profile with the highest personal utility [21-24]. Based on the choices respondents make, the impact of each attribute on the total utility is estimated, and the relative importance of the included attributes can be inferred [24-26]. The DCE developed as part of this study followed the guidelines of good research practice [27].

This study extends a previous DCE [15]. Recruitment and data collection in the previous study have been described [15]. The survey was extended to the following countries (in alphabetical order): Austria, Denmark, France, Germany, Ireland, Italy, the Netherlands, and Spain. Within each of these countries, a representative sample of 400 members of the general public was used. Respondents were recruited via a recruitment service, SurveyEngine [28], and received compensation according to customary agreements between SurveyEngine and the participant. Each national sample was representative (in composition according to gender and age) of the general population of each country. In order to achieve that, we provided the recruiting company with percentage data about the sample composition for age (18-39 years old, 40-59 years old, and ≥60 years old) and gender (female and male; % for each age group) based on the most recent (at the time of sample design) national statistics data available for each country. With these data, the recruiting company recruited participants of different ages and genders to obtain a sample with a defined number of respondents (ie, 400 respondents for the survey and 40 respondents for the ranking exercise) for meeting the requested composition. Data were collected from August to November 2020 for the original study, and from April to August 2021 for the other 8 countries.

This study was conducted in accordance with national and international laws and regulations regarding the protection of personal information, privacy, and human rights. Before starting the survey, prospective participants were informed about research participation (study aim, possibility of withdrawing from survey participation, possibility of withdrawing consent to the use of data for research until the data are analyzed, risks and benefits, and data processing). Participation was voluntary. Respondents provided consent to the use of the collected data for research purposes. Data collected by the recruitment service SurveyEngine were encoded. For data analysis, researchers accessed deidentified data.

Ethics approvals for the original study have been described previously [15]. For the extension of the study, the Ethics Committee of the Erasmus School of Health Policy and Management from the University (reference number 21-011) waived the necessity of formal testing by a medical ethics committee.

A 4-step approach was used for this process: (1) literature review; (2) 14 focus group discussions (including a nominal group technique [29] with members of the general population in the United Kingdom, Sweden, and Iceland); (3) expert interviews; and (4) confirmation ranking surveys. Steps 1 to 3 were part of a previous study described in detail elsewhere [15,30]. As a fourth step, an online ranking exercise was formulated and conducted in each of the additional 8 countries of interest to confirm the selected attributes and levels for the DCE. Respondents ranked a set of 12 items, which included the attributes and levels from step 2 listed above. This online ranking survey was sent to a representative sample of 40 respondents in each country. Recruitment and data collection for the ranking exercise were conducted from April to August 2021 by SurveyEngine. In all countries, the attributes included in the previous studies were among the highest ranked attributes across countries. The final attributes and levels are described in Table 1.

A Bayesian D-efficient design using 500 Halton draws and 1000 repetitions was constructed for this DCE, which was developed using NGene (version 1.2.1; ChoiceMetrics). Best guess estimates were used as prior information for the initial design. No interactions between attributes were included, and level balance was optimized. A total of 32 unique choice tasks were generated, which were divided over 4 blocks of 8 choice tasks. Respondents were randomized to a block and 2 situations of “type of information” (see below). Each choice task consisted of 1 profile representing a data sharing situation, and respondents could either accept or reject their data being shared in such a situation (Figure 1).

The online survey had 3 sections. The first part contained questions regarding demographic characteristics (eg, age, gender, nationality, and educational level), as well as questions on health-related professions, previous participation in research, use of digital technologies for health, and health status. The second part included the DCE. Each respondent was confronted with 8 choice tasks from 2 types of health information (ie, lifestyle information, physical health, mental health, or genetic information). This resulted in 16 choice tasks for each respondent. Prior to answering the choice tasks, respondents received detailed information on the type of health information that was applicable to the choice task and the meaning of all attributes and levels. Next, respondents were presented with an example choice task. The third part of the questionnaire measured trust and included related attitudinal questions.

The survey was further developed with 8 think-aloud interviews with members of the general public in the United Kingdom, Sweden, and Iceland, and a 2-day workshop with external experts (areas of expertise included law, philosophy, ethics, social science, and stated preference research) [15,30]. The survey was pilot tested in all countries (n=50 in the United Kingdom, Iceland, Sweden, and Norway, and n=40 in the other countries). Data gathered from the pilots of the previous study were analyzed, and attribute level estimates were used as prior input for the final DCE design. Pilot tests for this study were performed approximately 2 weeks before the data collection on a large scale in each country.

The survey and the ranking exercise were translated from English into the respective national language of each country included in the study.

For the statistical analysis, the countries were categorized into regions within Europe according to the United Nations classification. Northern Europe includes Sweden, Norway, the United Kingdom, Iceland, Ireland, and Denmark. Central Europe includes the Western European countries of the Netherlands, Germany, France, and Austria. Southern Europe includes Italy and Spain.

Descriptive statistics (means and frequencies) were presented for each country and region. Chi-square tests and 1-way ANOVA were conducted to examine differences in demographics across regions and across countries within those regions. Results were considered statistically significant at P<.05.

Panel latent class models (LCMs) were applied to determine attribute level estimates [31]. Separate models were conducted on the full pooled data set and within each of the 3 European regions considered. Such models account for the multilevel structure of the data and detect preference heterogeneity [32]. All attributes were considered nonlinear and were therefore effect-coded [24,33]. Based on model fit tests (Akaike information criteria and log likelihood), the model most suitable for our data was selected (models ranging from 1 to 6 classes were tested) [34]. The final utility equation is as follows:

A total of 5321 completed surveys were obtained. Due to a short completion time (less than 5 minutes), 306 completed surveys were excluded, and thus, 5015 completed surveys were included in the analyses (see Table 2 for a full overview of the demographics). The mean age of respondents across countries was 49.75 years.

The mean age and educational level of respondents differed significantly among regions. The highest mean age was in Central Europe. In comparison with Central and Southern Europe, there was a greater percentage of respondents with higher education in Northern Europe.

A 4-class LCM was fitted (Table 3). While class 1 respondents strongly a priori preferred not to share their health data, class 3 respondents were very positive toward data sharing. Class 2 and 4 respondents a priori preferred not to share their health data, but their preference could be heavily impacted by attribute levels of the situation. Respondents generally showed a disutility for a technological company collecting their data over their health care provider or an academic research project. They also, on average (except for class 2), preferred a national authority or academic research project to use their data over a technological or pharmaceutical company. Across all classes, respondents reported a strong disutility for using their data for promotion, advertising, and marketing over developing a new product or service, investigating a policy initiative, and evaluating quality. Respondents preferred being informed and asked to provide their consent over being informed and having the possibility to opt-out, only informed, and not informed. Respondents preferred the review of data transfer and use, or the review of the transfer over the absence of a review.

The most important attribute for the whole sample was “information on data sharing and consent,” followed by “availability of review process,” “reason for data use,” “data user,” and “data collector.” Moreover, “information on data sharing and consent“ was the most important attribute in all latent classes, except in class 4, where “reason for data use” was the most important attribute. While the other attributes seemed relatively unimportant for respondents in class 2, the opposite was found for respondents in class 4 (Figure 2).

As a data collector, class 1 and 2 respondents preferred an academic research project over their health care provider (not significant in class 1). The opposite pattern was found in classes 3 and 4. In all the classes, respondents expressed disutility for a technological company compared to their health care provider.

Compared to sharing data with a national authority, class 1, 3, and 4 respondents preferred to share data with an academic research project and showed disutility for a technological company. Class 3 and 4 respondents also showed disutility for a pharmaceutical company compared to a national authority, even though this was more acceptable than a technological company. In class 2, this attribute did not impact decision-making.

Health data sharing for promotion, advertising, and marketing purposes provided the most disutility in all classes. For class 1 respondents, developing a new product or service was preferred over quality evaluation. Sharing digital health data for quality evaluation was most acceptable by class 2, 3, and 4 respondents. For class 2 respondents, this was followed by investigating a policy initiative, while developing a new product showed disutility. For class 3 and 4 respondents, developing a new product or service was less acceptable than quality evaluation. For class 4 respondents, investigating a policy initiative was also acceptable but less than quality evaluation.

Furthermore, respondents preferred being informed and providing consent for data sharing over being informed and being offered to opt-out. Class 2 respondents also expressed disutility for being only informed. Sharing health data without being informed resulted in substantial disutility in all classes.

Class 1 and 3 respondents preferred review of data transfer only compared to review of data transfer and use. The opposite was found for class 2 and 4 respondents. The absence of review provided disutility in all the classes.

Compared to respondents from Northern Europe, respondents from Central Europe were more likely to belong to class 3 and less likely to belong to class 1 (compared to class 4). Respondents from Southern Europe were more likely to belong to class 2 and less likely to belong to class 3.

A 4-class latent model was fitted to the northern, central, and southern regions showing preference heterogeneity for sharing health data digitally within each region (see Tables S1-S3 in Multimedia Appendix 1).

In each regional sample considered as a whole, “information on data sharing and consent” was the most important attribute. This was followed by “availability of review process” or “reason for data use,” while “data user” or “data collector” was the least important attribute (Figure 3). Within each region, “information on data sharing and consent” was the most important attribute for most classes in Europe. Within each region, there was a class for which the other attributes were relatively unimportant compared to the most important attribute (class 3 for Northern Europe, class 2 for Central Europe, and class 2 for Southern Europe).

In Northern Europe, “information on data sharing and consent” was the most important attribute in all classes, except in class 2, where “reason for data use” was the most important, and “availability of review process” was relatively more important compared to the other classes.

Within Central Europe, a similar pattern was shown; however, for class 3, “information on data sharing and consent” and “reason for data use” were relatively equally important attributes. In 2 classes of Central Europe, “availability of review process” had less importance compared to “data collector” (for class 3 respondents) or “data user” (for class 4 respondents).

Finally, in Southern Europe, “information on data sharing and consent” was the most important attribute in classes 2 and 4, while “reason for data use” was the most important attribute in classes 1 and 3, where “availability of review process” was less important than “data user” or “data collector.”

The results showed that people in the European countries sampled shared, to a certain degree, general commonalities regarding what is important to them about digital health data governance. However, they also showed different priorities and preferences depending on their region of residence.

By design, the sample for each country aimed to reflect the respective national age range and gender distribution of the adult population (as of the most recent official information available from the respective national institutes of statistics at the time of sample design). Our data showed a generally higher proportion of female respondents and slight differences in mean age and educational level among the regions and countries. Previous studies reported that people of different ages had different levels of trust, risk perception, privacy concerns, data sharing attitudes, and willingness to share data [39,40,53,54]. Associations between education level and data sharing attitude, education, and preference for a review of data access were found in previous studies [35,40]. Further analysis aiming to investigate the relationship between the above-described factors and expressed preferences for governance mechanisms would be relevant in understanding and characterizing the variety of preferences in the European population.

This study did not include any Eastern European countries. Due to time and project budget constraints, we could not include additional countries in the study. This is a limitation of the generalizability of the results to Europe. It would be very important to extend the study further and ensure that all the European regions are covered to inform policy accurately and minimize possible biases in the results.

As an expansion of this project, the survey may be distributed in other countries worldwide. This would allow obtaining further insights, which would be valuable to grasp differences and similarities in people’s preferences for the governance of digital health data taking into account geographical regions and contexts of data sharing. In this paper, we decided to group the countries according to the UN classification of European areas. The use of the UN geoscheme would facilitate comparison and generalization according to a shared and globally known scheme in case the study is expanded worldwide. From a conceptual perspective, within the qualitative phase of the project, England, Iceland, and Sweden were grouped because those countries shared “similarities in breaches of trust among the public regarding secondary uses of health data” [42]. Additionally, in the first round of quantitative analysis [15], including the United Kingdom, Iceland, Norway, and Sweden, we implicitly followed the UN scheme; therefore, we decided to repropose the same grouping in this study. All the data were collected through online surveys during the COVID-19 pandemic. It was reported that the pandemic impacted patient preferences for data sharing, resulting in increased comfort in personal health data sharing compared to the prepandemic time [55]. We may speculate that during the pandemic, the growth of digitalization in every aspect of life [56], the role of the internet and media in providing health information, the emergence of digital health technologies (contact tracing apps and approaches for digital medicine) [57-60], and the efforts in data sharing for research and public health purposes [61,62] may have impacted the respondents’ attitudes and views on digital health data sharing and the expressed preferences. As the study was designed since its conception as an online survey, we believe that the findings were not affected by the method of data collection.

This study, which explored public preferences in 12 European countries, showed the co-existence of overarching priorities (such as the importance of information and consent) and heterogeneous preferences for contexts of data sharing among and within European regions. This study has confirmed the previous findings of a study in Northern European countries [15], provided further nuances to the preferences of Northern European countries, and added the preferences of residents in Western and Southern Europe. It allowed us to understand the pattern of preferences for digital health data sharing in a much broader context and according to geographical regions. With these results, we were able to discuss the challenges of data-sharing harmonization initiatives within Europe. Based on these results, we believe that there is no “one size fits all” governance solution. Instead, an interactive and dynamic model of informed consent offering individual granular control over data sharing accompanied by oversight mechanisms may be a valuable compromise to provide people with the ability to control the secondary use of their health data and to address their preferences for data flow within different contexts. These preferences are contrary to some of the proposals contained within the EHDS. Although the EHDS proposed an independent review of the secondary use of data by a new public entity, a health data access body, the draft regulation on the EHDS removes any role for consent or individual control in the secondary use of data, and it is proposed that individuals will not have a right to be informed about the purpose and the entity that has accessed and used their data. Harmonization initiatives seeking to provide a common ground for cross-border digital health data sharing should be developed upon empirical evidence. Understanding public preferences for digital health data sharing is important for developing adequate answers in policy-making and ensuring that new initiatives are perceived to be trustworthy and operating in accordance with people’s expectations.