2.1. Models of digital transformation in public administration

Models of digital transformation and technology adoption in public administration can be divided into two categories: (a) structural models, and (b) behavioral models. In this section, we provide an overview of these models and critically engage with their limitations in analyzing the drivers in policy-making processes.

3.1. Institutional landscape

Belgium is a federal state where tax and social security policies are mostly controlled at the federal level. However, regional and local governments have some authority in specific tax areas. Nonetheless, the fight against tax and social security fraud is the responsibility of the federal government.

Several layers of actors in the Belgian institutional landscape are involved in tax fraud detection. The first layer comprises the Federal Public Service (FPS) Finance and its directorate generals who are the primary responsible public bodies in fraud detection concerning tax collection, customs, and excise. The second layer of public sector organizations includes the financial intelligence processing unit (CTIF/CFI), several services of the police forces in charge of corruption (OCRC/CDBC) and of major crimes (DJSOC), the national bank (that has a point of contact—PCC/CAP—to alert it about frauds), the financial services and markets authority (FSMA), the college of attorneys-general and the gaming commission that provide information about potential tax infractions to FPS Finance. The third layer of actors is composed of individuals, public, and private organizations such as lawyers, accountants, notaries, banks, major auditing firms, and judges, who are mandated to provide information about fraud they encounter to FPS Finance. Lastly, some transversal actors at the EU and national level that enforce specific tax policy provisions act as external stakeholders (technology providers, unions, business federations, and NGOs in charge of fair taxation and/or privacy rights).

Social security infringement detection is part of the federal social security policy. Unlike the taxation domain, the policy domain on social security provisions is rather fragmented. The FPS Social Security oversees part of the policy design, but most of the policy implementation, including the fight against social security infringements, is led by a multitude of public bodies called social security public institutions (SSPI). There are five distinct SSPIs that are responsible for the detection of social security infringements: employment (ONEM/RVA), healthcare insurance (INAMI/RIZIV), social security for contractual workers (ONSS/RSZ), and social security for self-employed workers (INASTI/RSVZ), and control of social legislation (FPS Employment). These five organizations work together under the coordination of the SIRS/SIOD, which is responsible for developing an overall policy vision in the fight against social fraud. Some peripheral actors outside the federal government support the digital and data-sharing processes in this policy domain, such as SMALS, a nonprofit that acts as the federal IT support service, and the Crossroads Bank for Social Security (CBSS), which organizes data-sharing among SSPI. Nongovernmental actors such as social partners, social secretariats, and mutual health funds also contribute to the organization and management of social security policies. Eventually, the EU institutions can influence fraud detection processes in some social policy areas (e.g., social dumping) through the creation of specific regulations.

In addition to these domain-specific actors, there are also overarching political and administrative actors involved in both policy domains. These include the federal government, the federal parliament, the data protection authority (DPA/APD), and the Court of Audit (see the Supplementary Material for a detailed overview of the institutional landscape).

3.2. The use of advanced analytics in fraud detection

Advanced analytics is a data analysis methodology that uses predictive modeling, machine learning algorithms, natural language processing, deep learning, business process automation, and other statistical methods to analyze business information from a variety of data sources (Hanna et al., Reference Hanna, Burns and Presslar2022). Unlike traditional descriptive analytics, advanced analytics applies automation and AI to produce far deeper behavioral insights and predictions from complex datasets.

Fraud detection is a primary policy domain that uses advanced analytics. Advanced analytics can be used to analyze financial transactions, gain operational efficiency in fraud detection, and become more effective at investigating unwarranted spending on a large scale (West, Reference West2021). A 2020 report for the Administrative Conference of the United States found that 45% of the 142 agencies surveyed were using AI and/or machine learning (Freeman Engstrom et al., Reference Freeman Engstrom, Ho, Sharkey and Cuellar2020). Similarly, the UK and the Netherlands are other countries that have developed AI-led solutions for processing social benefit claims (Booth, Reference Booth2019; Kleizen et al., Reference Kleizen, van Dooren, Verhoest, Tan and Crompvoets2022).

The annual report of the OECD Tax series shows that the importance of advanced analytics is growing among tax authorities where data-driven insight is used against tax fraud such as in smart compliance with risk management and compliance by design (OECD, 2021). Advanced analytics are used in automating repetitive processes, extracting key data, scanning tax reports, identifying tax evasion, identifying tax deductions and credits, forecasting the burden of tax and improving transparency, and fighting against corruption (Forbes, January 9, 2020). Through supervised or unsupervised machine learning it is possible to find patterns of an anomaly in a large amount of data for predictive analysis and smart auditing that can outperform traditional data mining techniques (Van Vlasselaer et al., Reference Van Vlasselaer, Eliassi-Rad, Akoglu, Snoeck and Baesens2017; De Roux et al., Reference De Roux, Perez, Moreno, del Pilar Villamil and Figueroa2018). For example, the British Connect System uses advanced analytics on big data to detect fraud (Maciejewski, Reference Maciejewski2017).

However, technological, organizational, behavioral, trust, and regulative challenges, along with traditional digital barriers, seemingly affect the adoption processes of advanced analytics among taxation and social security authorities. Black boxes, biases, and model drift are certain concerns against the use of advanced analytics (Busuioc, Reference Busuioc2021). In a black box recidivism model, it may be unclear whether the high predicted risk of committing further crimes is based on admissible criteria (e.g., prior record) or potentially unlawful and/or discriminatory criteria (e.g., ethnicity) (Sandvig et al., Reference Sandvig, Hamilton, Karahalios and Langbort2016; Chander, Reference Chander2017). Moreover, policy choices for the area of application can cause algorithmic biases. For instance, in the Dutch SyRi case, the policy choices of municipalities to use advanced analytics in specific neighborhoods have led to false positives and discriminatory policies (Meuwese, Reference Meuwese2020). Where such biases eventually become public, they may give rise to major trust breaches among groups that the model was biased against, or even wider society (Kleizen et al., Reference Kleizen, van Dooren, Verhoest, Tan and Crompvoets2022).

Moreover, where an algorithm is a true black box, it may be difficult even for developers to ascertain the predictors used by the algorithm. The inability to supply an explanation of advanced analytics not only strands the trust toward public sector organizations but also may undermine the principle of transparency (Ananny and Crawford, Reference Ananny and Crawford2018; Ahonen and Erkkilä, Reference Ahonen and Erkkilä2020). Research shows that user cognitive routes concerning fairness, accountability, transparency, and explainability (Shin, Reference Shin2020), privacy (Shin et al., Reference Shin, Lim, Ahmad and Ibahrine2022a), transparent fairness (Shin et al., Reference Shin, Zhong, Biocca and Azmat2022b), and perceived humanness (Shin, Reference Shin2022) of algorithms may influence the perceived benefits and the willingness to use advanced analytics in policy-making processes.

Regulations also play a role in the way advanced analytics can be used in public policy domains. The use of advanced analytics in fraud detection by the Belgian public administration requires compliance with personal data protection rules and administrative law principles. The GDPR’s purpose limitation and data minimization principles must be observed, and data subjects’ rights, such as the right to information, access, and erasure, must be respected. Additionally, the algorithms used must not be biased and must not entail discrimination. The administration should also ensure that citizens have the right to understand the administrative decision, which is linked to the explainability of the decisions (Tombal et al., Reference Tombal, Willem, Terwangne, Tan and Crompvoets2022).

4.1. Thematic analysis

To address the first research question, we conducted thematic analysis on the interview transcripts to identify the variables perceived to be influential in the use of advanced analytics in fraud detection (Braun and Clarke, Reference Braun and Clarke2006; Fereday and Muir-Cochrane, Reference Fereday and Muir-Cochrane2006). The codification processes followed different cyclical processes, where an initial set of codes generated from interview transcripts were used to identify patterns and interrelations among initial themes, which were later categorized under overarching constructs and associated with the drivers identified by the theoretical models. For each transcript, we produced a theme record that synthesized all the themes, corresponding quotations, and theme categories. Themes and categories conveying similar meanings and substance were regrouped. The thematization process was conducted using a combination of inductive and deductive coding. Inductive processes are used in open and axial coding to uncover factors and their interrelationships that influence technology adoption processes. The categorized interrelationships were used later to interpret the direction of relationships during the development of the structural self-interaction matrix (see Section 5.1). Deductive coding was used to categorize themes from the behavioral and structural theories, while inductive (open) coding was used to identify new unforeseen themes and categories. To improve the reliability of the analysis, a researcher who was not involved in the initial data collection level performed a separate thematic analysis on the interview transcripts, and the results were compared before the final selection of categories. Through thematic analysis, we identified 13 variables perceived to be influential in the adoption of advanced analytics in the fight against tax and social security fraud in Belgium. Table 1 gives the categorization of constructs (i.e., variables), their definition, and the elements identified through thematic analysis. The subthemes used for the categorization of elements can be found in the Supplementary Material.

Table 1. Categorization of variables

4.2. ISM-MICMAC

Interpretative structural modeling (ISM) is a methodical and collaborative approach that mathematically derives and analyzes contextual relationships among factors identified through expert opinion, enabling scholars to establish hierarchical levels of challenges (Warfield, Reference Warfield1974). ISM analysis entails developing a directed graph that configures the relationships hierarchically as interpreted by the scholars. MICMAC (Matrices d’Impacts Croises Multiplication Appliqué a un Classement) analysis complements ISM by classifying factors based on driving power and dependence power (Ahmad et al., Reference Ahmad, Tang, Qiu and Ahmad2019). ISM has been employed in e-government literature to explore critical success factors in e-service delivery (Lal and Haleem, Reference Lal and Haleem2009), citizen’s perceptions of e-government’s trustworthiness (Janssen et al., Reference Janssen, Rana, Slade and Dwivedi2018), challenges for implementing the Internet of Things (IoT) in smart cities (Janssen et al., Reference Janssen, Luthra, Mangla, Rana and Dwivedi2019). In this study, we used ISM and MICMAC to explore and describe the dependency and driving powers between the variables identified by the thematic analysis. Typically, ISM begins with a literature review to identify key variables before interpreting the direction of relationships. However, in this study, we have adopted a novel approach and used thematic analysis as a preceding step, allowing us to integrate the perception of Belgian stakeholders in identifying the variables that influence the technology adoption processes. Moreover, we have used extracts from interviews to interpret the direction of relationships between drivers, ensuring that the interpretation was not solely based on the research team’s interpretation but also supported by the substance of the interviews. The adaptation of ISM in this research is illustrated in Figure 1.

Figure 1. ISM flowchart.

5.1. Structural self-interaction matrix

In developing the SSIM, there are four possible ways to relate variables, represented by “V,” “A,” “X,” and “O” symbols (Hughes et al., Reference Hughes, Rana and Dwivedi2020). The SSIM presented in Table 2 outlines the relationships between variables, with rows and columns indicated by “i” and “j,” respectively. The symbols are interpreted as follows: V, Variable i influences variable j; A, Variable j influences variable i; X, Both variables i and j are influenced by each other; O, Variables i and j are not related to each other or do not influence each other.

Table 2. Structural self-interactional matrix

Note. 1, Technological maturity; 2, Perceived usefulness; 3, Capacities, skills, and competencies; 4, Management/operational systems; 5, Perceived risk; 6, Governance system; 7, Technical infrastructure; 8, Public values; 9, Trust; 10, Sociocultural elements; 11, Interoperability; 12, Policy priorities; 13, Regulations; VR[i/j], variable i/variable j; i, row; j, column; V, Variable i influences variable j; A, Variable j influences variable i; X, Both variables i and j are influenced by each other; O, Variables i and j are not related to each other or are not influenced each other.

The research team, consisting of eight researchers, individually and separately assessed the relationships between variables before making a final decision. The relationships among the variables are interpreted through the links disclosed by the thematic analysis. Specifically, if an underlying subtheme for a variable is related to another variable, this relationship is used to interpret the nature of the relationship. Table 2 presents the final decision taken by the research team on the nature of the relationship between each pair of variables.

5.2. Reachability matrix

The SSIM was first converted to an Initial Reachability Matrix (IRM), which was then converted to a Final Reachability Matrix (FRM). The IRM illustrates the relationships described by SSIM in a binary way, with the following substitution rules: [1] if the (i, j) entry in the SSIM is V, the (i, j) entry in the reachability matrix becomes 1 and the (j, i) entry becomes 0; [2] if the (i, j) entry in the SSIM is A, the (i, j) entry in the reachability matrix becomes 0 and the (j, i) entry becomes 1; [3] if the (i, j) entry in the SSIM is X, both the (i, j) entry and (j, i) entry in the reachability matrix become 1; [4] if the (i, j) entry in the SSIM is 0, both the (i, j) entry and (j, i) entry in the reachability matrix become 0.

Next, the IRM was converted into an FRM, which includes transitive relations. Transitive relations occur when a variable X influences variable Y, and variable Y influences Z, so variable X should also influence variable Z, even if no mutual relationship is interpreted between variables X and Z. In such cases, an initial “no relationship” (i.e., “0”) was recoded as “1.” The IRM and FRM can be found in the Supplementary Material.

The FRM also shows the driving and dependence power of each variable. The driving power of a variable is the total number of variables, including itself, that it may help achieve. On the other hand, dependence power is the total number of variables, including itself, which may help in achieving it. These driving powers and dependence powers are later used in the classification of variables as part of the MICMAC analysis.

5.3. Level partitions

The FRM is used to create reachability and antecedent sets for each of the variables in the matrix. The reachability set, R(P_i), for a particular variable, includes the variable itself and other variables that may help achieve it (i.e., the corresponding value is 1). Similarly, the antecedent set, A(P_i) consists of the variable itself and other elements that may help in achieving it. The variables for which the interaction of these sets, $ R(Pi)\cap A(Pi)=R(Pi) $ , matches the reachability set, are considered the top-level variables of the ISM hierarchy. Each iteration of the level partition matrix identifies the hierarchy of variables in achieving other variables. The top-level variables do not assist in achieving other variables above their hierarchy level. Once the top levels are identified, they are separated, and the same process is repeated. The iteration process continues until all levels of partitions are established. The level partition matrix of each iteration is provided in the Supplementary Material. Table 3 displays the canonical matrix, which shows the level of each variable, along with their driving and dependence powers.

Table 3. Canonical matrix

Note. 1, Technological maturity; 2, Perceived usefulness; 3, Capacities, skills, and competencies; 4, Management/operational systems; 5, Perceived risk; 6, Governance system; 7, Technical infrastructure; 8, Public values; 9, Trust; 10, Sociocultural elements; 11, Interoperability; 12, Policy priorities; 13, Regulations; VR, variable; DEP, Dependence power; DRP, Driving power.

5.4. MICMAC diagram and ISM modeling

The MICMAC diagram visually categorizes variables based on their driving and dependency powers along the x- and y-axis. It contains four categories: autonomous, driving, dependent, and linkage. Autonomous variables have weak dependency and driving powers, and they are mostly disconnected from the system. Driving variables have higher driving power and weak dependency power, and they determine the changes in other variables without necessarily being dependent on changes in other variables. Dependent variables have higher dependency power and weak driving power, and they vary with changes in other variables without necessarily affecting the changes in other variables. Linkage variables are the most influential in the system with higher dependency and driving power. The cut-off point for each quadrant is arbitrarily designated by the number of variables. Since there are 13 variables, the cut-off point is designated at 6,5 on both axes. Figure 2 shows the positions of variables in the four quadrants.

Figure 2. MICMAC diagram.

The MICMAC diagram reveals that all variables are categorized as linkage, indicating that all identified variables hold significant influence in the adoption of advanced analytics in fraud detection. This finding implies that any changes in these variables could potentially result in systemic changes and impact the adoption strategies of AI and advanced analytics in fraud detection.

In the last stage, we developed the ISM based on the canonical matrix, and the direction of relationships identified in SSIM. The ISM in Figure 3 shows the direct and indirect relationships among variables. The direction of relationships is shown with an arrow. The interpretation of the model goes from bottom to up, where the variables at lower levels shape the perceptions regarding the variables in the following levels. Accordingly, the trust variable (variable 9) is based at the bottom of the ISM hierarchy, implying that the introduction of advanced analytics in fraud detection should start with evaluating the perceptions of trust determinants. The second level of hierarchy includes the interoperability variable (variable 11). The third level includes the perceived usefulness (variable 2) of advanced analytics in fraud detection. The fourth level is composed of the interrelationships between socio-cultural elements (variable 10), governance conditions (variable 6), technical infrastructure (variable 7), and management/operational systems (variable 4) in place. At the highest level, we observe a series of interrelationships between policy priorities (variable 12), regulation (variable 13), perceived risk (variable 5), public values (variable 8), the maturity of technologies (variable 1), and the capacity conditions (variable 3). In the following section, we interpret the directions of these relationships in terms of policy formulation processes concerning the use of advanced analytics in public policy processes.

Figure 3. ISM model.