Is everything under control? An experimental study on how control over data influences trust in and support for major governmental data exchange projects

2.1Setting the stage: Digital governance, control over data and transparency

Digitalization challenges typical notions of citizen-state interactions and interactions between public organizations (Busuioc, 2020). Effective data collection, storage and data-sharing processes can improve e-government applications (e.g. prefilled taxes and open data), predict citizen needs (e.g. employment training needs) and improve the efficiency of services (e.g. near real-time updates of dossier statuses). Simultaneously, digitalization in government raises questions on the degree to which public organizations should be entrusted with personal data, the reach of the right to a private life and the safety of data managed in governmental projects (e.g. Janssen & Van den Hoven, 2015).

The ethical concerns related to digitalizing governments have resulted in various approaches to make data collection, data sharing and data analytics processes more ‘human-centric’, i.e. putting users and/or stakeholders first in regulation, policy and project design. Such human-centric approaches frequently place a dual focus on transparency and control. For instance, the GDPR is based on the principle that data subjects must be informed in a clear and understandable way what data is processed for what purposes, before consent is given (Van Ooijen & Vrabec, 2019). Moreover, it codifies the right to erasure and right to access stored personal data, making user control increasingly relevant for both private and public sector entities. The more recent California Consumer Privacy Act (CCPA) provides for a similar consent- and rights-based approach, illustrating the glowing clout of control-based regulatory efforts in recent years. Similarly, soft law approaches such as the EU’s and UN’s trustworthy AI guidelines consistently incorporate autonomy, user control and transparency as core principles, suggesting that control and informed consent are internationally becoming important ethical principles (Floridi, 2019; Kozyreva et al., 2020).

Recent technical developments such as Self-Sovereign Identity (SSI) systems take the idea of building trust through transparency and control even further (Naik & Jenkins, 2020; Flanders Investment & Trade, 2022). Such systems would provide citizens with an online ‘digital wallet’, in which they can see what data is available on them and who may access it. For some data categories, citizens might even exercise some degree of control, determining whether to share such data with certain actors or not (Naik & Jenkins, 2020). SSI projects are still in an embryonic state, their emergence and simultaneous developments in the regulatory and policy arena’s (the aforementioned GDPR and guidelines). Nonetheless, their trust-building aim raises the question how control over data psychologically relates to citizen attitudes towards data-driven government projects. Although a new development, digital wallets are no longer entirely theoretical either, with e.g. a Belgian consortium working on an experimental SSI system with an associated digital wallet (Olsthoorn, 2022).

Taking inspiration from both regulatory efforts and technical developments in e.g. the SSI domain, we examine how information on control and actual control would empirically affect citizen attitudes. In the remainder of this section we will therefore first discuss three citizen attitudes that could be affected by information and control (trust in project, policy support and policy concerns). Subsequently, we discuss how information, transparency and control are causally linked with trust, policy support and policy concerns. Finally, we discuss how control over data may affect trust, policy support and policy concerns.

2.2Trust in project, policy support and policy concerns

Citizens possess multiple analytically distinct attitudes towards governmental projects, including public authorities’ data processes. In this study, we examine three different, albeit related attitudes towards data exchange projects in government: trust in (data exchange) project, policy support and policy concerns. Trust in data exchange project is seen as a manifestation of trust in government where a specific project – instead of the government or a public authority as a while – is the trust referent. In a public sector setting, trust is usually defined as the expectation of a trustor that a government will engage in opportunistic behavior that results in harm to the trustor (Hamm et al., 2019; Grimmelikhuijsen & Knies, 2017). We argue that our notion of trust in project is more aligned with trust in government than with trust in technology or trust in systems (see Lankton & McKnight, 2011 for an overview), as citizens will not just evaluate the reliability and functionality of the data exchange system itself, but also the actions, characteristics and intentions of the civil servants and public authorities developing, managing and using it. In our case such evaluations may e.g. also include follow-up decisions by public authorities based on exchanged data, making the trust referent the broader project, rather than ‘just’ the technological system.

Trust is conceptualized as multidimensional in nature, with social scientists commonly distinguishing between the dimensions ability (a trustee’s competence to act in a way beneficial to the trustor), benevolence (the degree to which the trustee takes the trustor’s best interests to heart) and integrity (the degree to which a trustee follows the rule of the game when interacting with the trustor) (O’Hara, 2012; Degli Esposti et al., 2020; Hamm et al., 2019; Grimmelikhuijsen & Knies, 2017). Trustors by definition operate under incomplete information on the behavior and intentions of a trustee, and therefore have to rely on incompletely informed assessments of a trustee’s trustworthiness (Hamm et al., 2019). Faced with limited but gradually updating information, people may change their beliefs regarding the trustworthiness of a trustee – in this case a data exchange project – based on novel information on any of the three trust dimensions (Degli Esposti et al., 2020).

Although important, trust is not the only relevant citizen attitude. One could for instance trust that a government project has the best intentions and will be performed capably and with integrity, without supporting the content of that project. To gain a more comprehensive view of citizen attitudes, we also examine two slightly different – albeit likely correlated – concepts. Policy support captures the degree to which citizens are supportive of data exchange projects in government (Zahran et al., 2006). Slightly different is the policy concerns concept, which concerns the degree to which respondents have reservations on specific aspects of a policy or public sector project. A significant impact of our treatments on either variable would suggest citizens have updated their cognitive assessment of data exchange projects. The following subsections now turn to linking information on control and the exercise of control to trust in project, policy support and policy concerns.

2.3Information on control over data and its impact on citizen attitudes

Although digital government has often been connected with opportunities for increased transparency and – through such transparency – greater citizen trust and accountability (Matheus & Janssen, 2020; Bertot et al., 2010), substantial challenges remain. Previously, we already noted that information on, and control over the way governments process data is becoming increasingly relevant in practice. However, the question remains how governments should provide information and control to citizens, and what transparency exactly means in a digital governance context. More passive, descriptive definitions of transparency emphasize that a transparent government bestows the ability to see through the windows of an institution (Meijer, 2009). Others see transparency as a principle that reinforces other processes. Building on this perspective, Buijze (2013) argues that transparency holds two important functions. First, it facilitates citizen and stakeholder decision-making. By providing citizens with information on what public authorities are doing, citizens become empowered to determine their own opinions and act accordingly. Second, transparency enables outside scrutiny of public authorities’ behavior. Information provided in line with the second function allows citizens to hold governments to account and serves to protect the rights of citizens (Buijze, 2014; see for a similar perspective, Stirton & Lodge, 2001).

The approach provided by Buijze (2013) is useful, as it also implies that information does not necessarily equate to transparency – and in some cases may even adversely affect it. For information provided by a government to contribute to transparency, it should contribute to citizen will-formation, to the monitoring of government actions or to the protection of citizens’ positions. At the same time, excessive or incomplete information may in some cases adversely affect citizen decision-making or the scrutiny of public authorities (Buijze, 2013). This is highly relevant for a digital governance context. As noted by O’Hara (2012), data processes in governments are regularly so complicated (both in technicality and in scope) that any attempt at being transparent on them entails difficult choices what information should be provided and not. Such choices are presented both in terms of the information that is presented to citizens and in terms of the design of the context in which this information is presented (e.g. the website on which it is placed, or regarding the other information sources available on that same website).

Previous work has observed that this design process may have important – and sometimes unexpected – consequences for citizen attitudes, with different forms of presenting information yielding varying levels of trust and privacy concerns (e.g. Brandimarte et al., 2013; Van Ooijen & Vrabec, 2019; Brandimarte et al., 2013). Van Ooijen & Vrabec (2019) e.g. note that even the mere presence of information such as a privacy policy may suggest to users that data may not be shared by the processor, even though the privacy policy in question does not strictly speaking incorporate provisions to that end. Such results suggest that the signals provided by information on websites an important determinant of the concerns and trust that people have regarding data processes (Berliner Senderey et al., 2020), and that some information may even allow for a degree of manipulation of citizen perceptions (Brandimarte et al., 2013; Alon-Barkat, 2020). Moreover, they imply that website design may even provide room for manipulation by the creators of websites. More optimistic on the empowering and trust-building effects of information is the broader public administration literature on transparency and data processes. Matheus and Janssen (2020) note in their recent review that Open Government Data (OGD) is expected by a majority of contributions to have positive effects on citizen trust. Moreover, research on how users respond to the use of AI in government has suggested that explainable models, i.e. systems providing information on how predictions are generated, are generally perceived more favorably by users than non-explainable counterparts (Grimmelikhuijsen, 2022), suggesting ‘algorithmic transparency’ on how data processes work may have positive effects. Aoki (2020) similarly demonstrates that communicating legitimate purposes of chatbots used in governments may enhance public trust. Notable is that such studies focus more on the content of what is communicated, whereas studies such as Brandimarte et al. (2013) and Van Ooijen and Vrabec (2019) focus more on potentially misleading ways information is presented. This again reinforces the idea that transparency and information do not equate to one another, and that in some cases information may even adversely affect transparency (Buijze, 2013). Nonetheless, in both cases, studies seem to show that positive information (whether genuinely transparent or incomplete or misleading) allowed entities to improve attitudes such as trust in and support for their data processes.

In sum, both the literature on privacy options in websites and the public administration literature on transparency in data processes seems to suggest that providing explicit information on the control that citizens would have over their data would have positive effects on attitudes such as trust and support – even if the underlying intentions sometimes seem to differ. We will integrate these insights into our experiment by providing additional information on the control that citizens will have over their data in the context of a novel data exchange platform, and expect that more information on control should result in greater trust and policy support, while reducing citizen concerns on data exchange projects.

2.4The impact of exercising control on citizen attitudes

Beyond studies on the effects of information and transparency, the impact on control over data has also increasingly received scholarly attention. Prior studies have generally shown that increases in (perceived) control positively affect citizen attitudes. Contributions in this line argue that data collection procedures are often opaque in nature (Malhotra et al., 2004), requiring citizens to place a certain amount of trust in the data processor that data will not be misused. Such misuse might come in various forms, such as the use of data for purposes that are not considered desirable or the recombination of data against the wishes of the person providing the data (Whitley, 2009; Bilogrevic & Ortlieb, 2016; Malhotra et al., 2004). What is more, considering that data is often stored long-term, this brings about further unease on data storage terms and the potential for future misuse (Van Ooijen & Vrabec, 2019). Essential here is that attitudes and behavior will flow from perceived misuse, which does not necessarily align with what is legally possible with the data. For instance, long-term storage may be perfectly legal under some privacy policies agreed to by someone, but may – due to bounded rationality and incomplete information – still be perceived as misuse.

Such fears of misuse may be ameliorated by granting citizens some measure of (perceived) control. Here, it is important to note that control and perceived control are subtly but importantly different. In the context of Facebook, Brandimarte (2013) for instance shows that awareness of control mechanisms to limit whether media can be shared with friends and/or strangers gives respondents a higher level of perceived control, even though actual control over the degree to which personal information is sold to advertisers stays limited. Perceived control may function to reduce uncertainty and unease regarding the potential consequences of sharing data, as the person involved now feels directly empowered to generate a desirable outcome (Whitley, 2009; Van Dyke et al., 2007). Therefore, being provided with a measure of actual control by the data processor may – through improved perceived control – function as a sign of trustworthiness and procedural justice, heightening the degree to which the person involved experiences is willing to provide data to the party in question (Malhotra et al., 2004). Moreover, the role of (perceived) control will likely become stronger as the (perceived) opportunity for abuse becomes higher (Malhotra et al., 2004). Public sector organizations tend to have greater access to personal data than their private counterparts, as well as incentive to share sensitive categories of information (such as data on financial position, criminal convictions, migration status, or previous benefits received). As such, adjusting levels of control over data collection may be a particularly successful predictor of trust in project, policy support and concerns in the context of public sector data projects.

Being granted immediate control over the data categories shared may trigger a process in which respondents cognitively focus on the trustworthiness of the interface directly presented to them – potentially to the detriment of other facets of a governmental projects’ trustworthiness (Brandimarte et al., 2013). Brandimarte et al. (2013) argue that this may even lead to unintended oversharing of data, as limited cognitive capacity (bounded rationality) leads humans to focus on the immediate control they have over data at the expense of what may happen to that data in the future (e.g. during the processing and analysis stages after data is exchanged between governments), leading to reduced perceived risk that data will be abused. This should be visible in our results, as the reduced levels of perceived risk associated with the presented data exchange platform may in turn illicit heightened trust in project and policy support (Brandimarte et al., 2013; Alon-Barkat, 2020), while reducing respondents’ policy concerns.

In sum, we expect that increased control will increase trust and support in data exchange platforms, while reducing policy concerns regarding such platforms.

3.1Experimental design and treatments

The experiment relies on a group structure in which groups differ on one factor. Three groups are given a short common baseline to which three different additions are made. The common baseline shows information on a hypothetical data exchange project, inspired on existing systems in Belgian governments. More specifically, Belgian governments host platforms such as MAGDA (Flemish regional government (European Commission, n.a.)) and the Crossroads banks (federal government), to which public entities can connect to request data that is stored in a decentralized fashion across multiple (mostly public, but sometimes private) entities (CBSS, n.a.). The commonality between these systems is that they seek to provide efficient, safe and legally compliant platforms for data exchange between public (and sometimes private) organizations. At the same time, many of the data sources that can be connected concern personal and/or sensitive data categories, which may raise concerns among some citizens. This makes data exchange platforms an interesting (and underexamined) context to examine how variations in transparency and control over data may affect attitudes.

The experiment introduces two treatments. Two groups are provided with one treatment, while a third receives a combination of both treatments. The first treatment consists of an exercise where respondents can indicate what data categories they would like to share with various private and public parties. This exercise serves as a direct signal to citizens that they will be able to exercise some degree of control over their data (Brandimarte et al., 2013). The 8 listed data categories were chosen to correspond to both the data categories registered with the Belgian Crossroads bank for social affairs and the data categories that could be shared between governments under the Dutch System Risk Indication (SyRI). This prevents the intervention from being relevant to only one setting. Data categories include identification data, labour data (including career and wage), welfare data (e.g. benefits received), data on prior interactions with the criminal justice system, real estate data, education status and registrations of vehicles. One category incorporated in SyRI but excluded here is migration, as this category may not be relevant for various respondents. In the second treatment, information on a hypothetical governmental project that would improve citizens’ control over their data is provided. The treatment is provided through a brief paragraph of textual information, and is inspired on advances in digital wallet and SSI systems (such as the Belgian SOLID project (Olsthoorn, 2022)), through which citizens would be able to exercise some measure of control over what data is shared with what actor (Naik & Jenkins, 2020).

Figure 1.

Screen 1, showing baseline and (for full intervention group and information on control group) the information on control intervention.

Figure 2.

Screen 2, version 1, showing a list of data types shared between governments to respondents that do not receive the direct control treatment.

The group receiving the full treatment receives both the exercise on direct control and the textual information on a new project to improve citizen control. This group thus receives the strongest intervention. The group receiving the full intervention is designated as the reference category in regression analysis, as the other two groups each differ on one component with the group receiving the full treatment. Figures 1–3 show the original Dutch screens incorporating the experimental treatments, with Fig. 1 containing the baseline and information on control intervention, while Figs 2 and 3, respectively contain the screen shown to the group of respondents not receiving the direct control and the screen shown to respondents in the group that do receive the direct control. Translated vignettes are available in Appendix 1. A schematic overview of the information available to respondents in the three groups is available in Table 1.

Figure 3.

Screen 2, version 2, showing the direct control treatment (the same list of data types as version 1, but now with the possibility to indicate with whom a respondent would like to share particular data types).

Respondents are randomly assigned to groups, using simple randomization and Qualtrics software to automatize the process, ensuring that unobserved characteristics are distributed randomly across observations. Groups are equal in size (∼ 125 respondents).

3.2Sample size, recruitment and pretesting

The survey containing the experiment was held through a marketing company’s panel of Flemish (Dutch-speaking region of Belgium), adult participants. The panel approaches its members until the intended amount of respondents is reached. Although recruitment into the panel is done through self-selection, selection of potential recruits is done so that a representative sample of the Flemish population (based on age and gender and within a weighting factor of 3) is reached. Panel members receive payment through points per survey completed, which they can subsequently spend on offers made by the marketing company. [Anonymized university] is not involved in this payment scheme. Final sample size slightly exceeded our aims, consisting of 402 observations. After random assignment, the full treatment group has 133 observations, the information on control group has 134 observations and the direct control group has 135 observations. This exceeds the aim of our power analyses, which suggested we need just over 100 participants (see Appendix 2).

We held pretest using the cognitive interview (read-out-loud) protocol with 5 respondents. This protocol requires respondents to read out loud every component of the survey and mention any comments they have, for instance on sentence structure, complexity, jargon or issues with the interpretation of items. The benefit of this approach is that a great amount of insight is created on why certain items fail to produce the desired associations among respondents. Subsequently, a soft launch was held with the survey company to test for issues that result in heightened drop-out rates, which revealed no further issues.

3.3Measurements