129 2591-2259 / This is an open access article under the CC-BY-SA license https://creativecommons.org/licenses/by-sa/4.0/ Lazović Pita, L., Odžak, A., Šćeta, L. (2025). Examining Individual Tax Morale in Europe with Machine-Learning Methods. Central European Public Administration Review, 23(1), pp. 129–155 DOI: 10.17573/cepar.2025.1.05 1.01 Original scientific article Examining Individual Tax Morale in Europe with Machine-Learning Methods Lejla Lazović Pita University of Sarajevo, School of Economics and Business, Bosnia and Herzegovina lejla.lazovic@efsa.unsa.ba https://orcid.org/0000-0001-9421-1842 Almasa Odžak University of Sarajevo, Faculty of Science, Bosnia and Herzegovina almasa.odzak@pmf.unsa.ba https://orcid.org/0000-0001-6269-9759 Lamija Šćeta University of Sarajevo, School of Economics and Business, Bosnia and Herzegovina lamija.sceta@efsa.unsa.ba https://orcid.org/0000-0002-0410-0311 Received: 3. 12. 2024 Revised: 21. 2. 2025 Accepted: 24. 2. 2025 Published: 20. 5. 2025 ABSTRACT Purpose: This research examines and contributes to the behavioural lit- erature on voluntary tax compliance. It focuses on the use and potential of machine-learning (ML) methods and models to predict individual tax morale across Europe, and it identifies the factors that influence predic- tive accuracy. Design/Methodology/Approach: Using data from the fifth wave (2017– 2020) of the European Values Survey (EVS), a data-driven, systematic approach employing six ML methods is applied to predict individual tax morale across Europe. The importance of formal, informal and socio-de- mographic factors is assessed, and the study tests whether incorporating the Corruption Perception Index (CPI) improves predictive accuracy. Findings: The results indicate that ML methods and models can enhance understanding and prediction of individual tax morale in Europe. Among the deployed models, artificial neural networks (ANNs) achieved the highest accuracy. Accuracy increased across all ML methods when the CPI was included. Attitudes towards bribery, perceptions of immigrants’ im- Central European Public Administration Review, Vol. 23, No. 1/2025 130 Lejla Lazović Pita, Almasa Odžak, Lamija Šćeta pact on the national welfare system, and gender emerged as significant formal, informal and socio-demographic factors. Academic contribution to the field: The study offers a novel application of data-driven ML methods to the prediction of individual tax morale. Given the scarcity of empirical ML research in the social sciences, the find- ings provide valuable insights in a European context and may serve as a basis for further global research. Practical Implications: The conclusions are particularly relevant for gov- ernments and tax administrations seeking to improve tax compliance and revenue collection. In the European context, the results confirm the virtu- ous circle linking effective government performance, high tax morale and voluntary tax compliance—insights that are crucial for decision-makers, regulators, European institutions and tax-policy makers. Originality/Value: The findings confirm that, when ML methods are ap- plied, individual tax morale can be viewed as an outcome of interactions between formal and informal institutions. They also show that predictive accuracy is higher in countries with lower corruption, as indicated by a higher CPI. Keywords: corruption, EVS, individual tax morale, machine learning Preučevanje individualne davčne morale v Evropi z metodami strojnega učenja POVZETEK Namen: Raziskava proučuje in nadgrajuje vedenjsko literaturo o pros- tovoljnem izpolnjevanju davčnih obveznosti. Osredotoča se na uporabo in potencial metod ter modelov strojnega učenja za napoved individu- alne davčne morale po Evropi ter opredeljuje dejavnike, ki vplivajo na napovedno natančnost. Zasnova/metodologija/pristop: Na podlagi podatkov petega vala (2017– 2020) Evropske raziskave vrednot (EVS) je bil uporabljen podatkovno voden, sistematičen pristop, ki vključuje šest metod strojnega učenja za napovedovanje individualne davčne morale v Evropi. Ocenjena je bila pomembnost formalnih, neformalnih in soci-demografskih dejavnikov, hkrati pa je študija preverila, ali vključitev Indeksa zaznave korupcije (Cor- ruption Perception Index – CPI) poveča napovedno natančnost. Ugotovitve: Rezultati kažejo, da lahko metode in modeli strojnega učenja izboljšajo razumevanje ter napovedovanje individualne davčne morale v Evropi. Med uporabljenimi modeli so umetne nevronske mreže dosegle najvišjo natančnost. Pri vseh metodah strojnega učenja se je natančnost povečala, ko je bil v model vključen CPI. Odnos do podkupovanja, zaznave vpliva priseljencev na nacionalni socialni sistem in spol so se izkazali za pomembne formalne, neformalne in soci-demografske dejavnike. Akademski prispevek k področju: Študija je nov, podatkovno usmerjen pristop uporabe metod strojnega učenja za napovedovanje individualne davčne morale. Zaradi redkosti empiričnih raziskav strojnega učenja v družboslovju ugotovitve ponujajo dragocene vpoglede v evropskem kon- tekstu in so lahko temelj za nadaljnje globalne raziskave. Central European Public Administration Review, Vol. 23, No. 1/2025 131 Examining Individual Tax Morale in Europe with Machine-Learning Methods Praktična uporabnost: Sklepi so posebej pomembni za vlade in davčne uprave, ki želijo izboljšati davčno skladnost in stopnjo pobiranja prihod- kov. V evropskem kontekstu rezultati potrjujejo krog pozitivnih povrat- nih zank, ki povezuje učinkovito delovanje vlade, visoko davčno moralo in prostovoljno izpolnjevanje davčnih obveznosti. Gre za spoznanja, ki so ključna za odločevalce, regulatorje, evropske institucije in oblikovalce davčne politike. Izvirnost/vrednost: Ugotovitve potrjujejo, da lahko pri uporabi metod strojnega učenja individualno davčno moralo obravnavamo kot rezultat interakcij med formalnimi in neformalnimi institucijami. Prav tako kažejo, da je napovedna natančnost večja v državah z nižjo stopnjo korupcije, kar se kaže v višjem CPI. Ključne besede: korupcija, EVS, individualna davčna morala, strojno učenje JEL: H26 1 Introduction In recent years there has been a growing academic interest into investiga- tion of values, attitudes and social norms as possible explanations to human and economic behaviour, particularly within the extensive body of tax com- pliance literature (Alm and Torgler, 2007, p. 1). Even though some research into tax morale has been conducted in the late 1960s by Strümpel (1969) and Schmölders (1970; Alm and Torgler, 2007, p. 4), the empirical investigation of the tax morale concept began with the World Values Survey and the Europe- an Values Survey (WVS and EVS respectively) in the 1980s. Hence, the article wishes to contribute towards the understanding of human and economic be- haviour in the investigation of individual tax morale across Europe by applying ML methods and models for the first time. In academic literature and empirical research, several terms are closely asso- ciated with the concept of tax morale, including tax evasion and tax compli- ance. In theory, the social psychology has set the question about the relation- ship between tax evasion and tax morale (Lewis, 1982, p. 165), while empirical studies have provided strong evidence of a negative correlation between tax morale and both – tax evasion and – the size of the shadow economy (Tor- gler, 2005, p. 135; Sá et al, 2013, p. 1). Empirical studies have mostly relied on either WVS or EVS datasets. Furthermore, authors have investigated the concepts of tax morale and tax evasion separately, whereby tax evasion has been the subject of more than 40 years of academic research, including stud- ies on tax evasion games (Alm and Malézieux, 2021, p. 699). In the context of the European Union (EU) and its member states, research on tax evasion has been growing, especially in relation to the size of the shadow economy (Sch- neider, Raczkowski and Mróz, 2015, p. 34) and an estimation of substantial loss of tax revenues due to Value Added Tax (VAT) frauds (Dobrovič, Rajnoha, and Šuleř, 2021, p. 705). In addition, significant differences exist among EU member states, with newer members (those that joined post-2004) experi- Central European Public Administration Review, Vol. 23, No. 1/2025 132 Lejla Lazović Pita, Almasa Odžak, Lamija Šćeta encing higher levels of tax evasion compared to older member states (Yamen et al., 2018, p. 26; Dobrovič, Rajnoha, and Šuleř, 2021, p. 705). For example, in 2019, the EU average share of tax evasion expressed as a share of total tax liability amounted to 20 percent, with great differences among the EU mem- ber states (e.g. tax evasion in Sweden as an old EU member amounted to only 2% and Romania as a new EU member to 49% (Dobrovič, Rajnoha and Šuleř, 2021, p. 722)). Due to the negative consequences of tax evasion on public revenues and the tax system equity (Sá et al, 2013, p. 1), the EU institutions are actively engaged in researching and prevention of tax evasion, but also on methods and ways to increase tax morale. For instance, the EU Tax Observa- tory publishes global tax evasion reports, or organizations such as OECD have most recently reported on tax morale (OECD, 2019, p. 4). The investigation in concepts related to tax morale as a component of tax compliance (Luttmer and Sighal, 2014, p. 150) has its roots in the Alling- ham and Sandmo (1972, pp. 323–324) benchmark model of tax evasion. So, the academic interest and related research to why people pay taxes (or do not) has been examined throughout decades, and yet still the understand- ing remains limited (Dulleck et al., 2016, p. 9). Authors deal with this issue from different perspectives all grounded in the traditional economics-of- crime approach (Becker, 1968, p. 173; Dulleck et al., 2016, pp. 9–10) which is elaborated through tax compliance. Tax compliance could be viewed from three perspectives: general deterrence theory, economic deterrence models and fiscal psychology (Riahi-Belkaoui, 2004, p. 137). Tax compliance or non- compliance (as investigated through the concept of tax evasion) have been evaluated in all three concepts and their economic and political impact in the society have also been measured (Riahi-Belkaoui, 2004; Barone and Mocetti, 2011, p. 5). However, the social psychology question to why people pay or evade taxes (Lewis, 1982, p. 165), still remains open, so, in order to solve a puzzle, researchers have introduced a concept of ‘tax morale’ as an ‘individual intrinsic motivation to pay taxes’ (Feld and Frey, 2002, p. 88; Riahi-Belkaoui, 2004, p. 137; Barone and Mocetti, 2011, p. 5). Due to the negative correlation between tax morale and tax evasion, the literature dealing with tax evasion is as comprehensive as the literature on tax morale and it has been investi- gated and empirically researched through the concept of ‘tax evasion games’ using different methods and models for more than 40 years now (Alm and Malézieux, 2021, p. 700). In this article, the focus is on applying new methods and models in the prediction of individual tax morale which contributes to- wards tax compliance literature. Tax morale, as an important and popular concept that increases voluntary tax compliance (Luttmer and Singhal, 2014, p. 151) has been gaining prominence in the empirical research over the past three decades. Different aspects of tax morale have been investigated so far – from five possible tax morale mecha- nisms (intrinsic motivation, reciprocity, peer effects and social influences, cul- tural factors and information imperfections and deviations from utility maxi- mization) defined in the works of Luttmer and Singhal (2014, pp. 155–163), to more specific investigation of determinants of tax morale (Hofmann et al., Central European Public Administration Review, Vol. 23, No. 1/2025 133 Examining Individual Tax Morale in Europe with Machine-Learning Methods 2017, pp. 64–65). In terms of factors that influence tax morale, three main categories have been recognized (formal and informal institutions and socio- demographic characteristics, Horodnic, 2018, p. 870). The OECD highlights the significance of the virtuous circle between effective government performance and high tax morale/high tax compliance (OECD, 2018; 2019) indicating the significance of formal institutions as defined by Horodnic (2018, p. 871) or reciprocity as a tax morale mechanism defined by Luttmer and Singhal (2014, p. 157). The inevitable party of interest is the tax administration as the executive power of the state that keeps records and col- lects tax revenues. Understanding concepts and determinants are of a crucial importance to governments, tax administration officials and tax policy mak- ers, all aiming to improve tax compliance and revenue collection. The research uses ML methods in the prediction of individual tax morales since ML methods have not yet been used. The results of the research ought to help tax administration officials and tax policy makers in improved predic- tion accuracy regarding individual tax morale as well as the identification of key determinants of individual tax morale primarily across Europe. Further- more, results may be replicated to other countries as well. The significance of such interdisciplinary research being conducted for the first time has been highlighted in the works of Athey (2018, p. 507). So, for the first time, present research wishes to contribute to current theoretical and empirical literature gap related to the application of ML methods and models in the prediction of individual tax morale. After the introduction section, the paper is divided into four parts: a brief literature review, research design and methodology, results and discussion and conclusion sections. 2 Literature Review The investigation of tax morale takes an interdisciplinary approach, drawing from multiple scientific fields all trying to explain human and economic be- haviour. Following the arising questions in the traditional economics-of-crime approach (Becker, 1968, p. 173; Allingham and Sandmo, 1972; Dulleck et al., 2016, pp. 9–10) related to why people pay (or not) taxes, the concept of tax compliance has been gaining academic prominence over last few decades (Luttmer and Singhal, 2014, p. 166). However, the basic economics-of-crime approach cannot fully capture the vibrant nature of decisions and motivations to comply or evade taxes as explained in tax compliance theory and practice (Dulleck et al., 2016). Dulleck et al. (2016, pp. 13–16) argue that other factors, social norms, individual, social, and cognitive dissonance affect tax compli- ance. Even though Luttmer and Singhal (2014, p. 151) consider enforcement as the primary driver of tax compliance, authors also argue that concept of tax morale is an important component of voluntary tax compliance. The concepts of tax morale and tax evasion have been theoretically intro- duced and investigated in the social psychology in the early works of Lewis (1982, p. 165). The empirical investigation and possible connection between Central European Public Administration Review, Vol. 23, No. 1/2025 134 Lejla Lazović Pita, Almasa Odžak, Lamija Šćeta attitudes and behaviour have been exhaustively investigated with the WVS and the EVS surveys introduced in the 1980s, and several questions thereof that might explain and measure the level of tax morale, or the size of tax eva- sion (Torgler, 2005, p. 136; Sá et al., 2013, p. 1). The tax compliance concept in economic theory is closely examined through concepts of tax avoidance and tax evasion as explained in seminal works of Allingham and Sandmo (1972, p. 323) and Yitzhaki (1974, p. 201). Over more than 50 years of study, researchers have tried to disentangle individual de- terminants of tax compliance. Alm et al. (1995, p. 6), have used experimental methods to explore the major factors – economic and noneconomic – that affect tax compliance: detection and punishment, the burden of taxation, public good provision, overweighting of low probabilities, and social norms. Authors such as García et al. (2018, p. 8) argue that the causes of tax evasion acceptance may be divided into three categories: internal or individual that de- pends on values, contextual, social or institutional that focuses on differences across countries and regions (Lago-Peñas and Lago-Peñas, 2010, p. 441) and a third category introduced by Alm (2014) that combines group motivations such as altruism and fairness (Alm, 2014, pp. 261–267; García et al., 2018, p. 8). Most recently, Horodnic (2018, p. 878) in the systematic literature review argues that tax morale could be perceived as an outcome of interaction be- tween formal and informal institutions, whereby socio-demographic variables are mostly used as control variables. Following the identification of five classes of tax morale mechanisms (intrinsic motivation, reciprocity, peer effects and social influences, cultural factors and information imperfections and deviations from utility maximization) defined in the works of Luttmer and Singhal (2014, pp. 155–163), the first mecha- nism of intrinsic motivation measures tax morale as a feeling of guilt or shame (Andreoni et al., 1998, pp. 850–852). With the EVS cross-country surveys in several waves, the academic research in the measurement of tax morale and its determinants through EVS questionnaire within and across countries has been exhaustive (Hofmann et al., 2017, p. 65). In all research, the depend- ent variable taken as a representative of the level of tax morale has been the statement from the EVS survey responded at the scale from 1 (never) to 10 (al- ways): Please tell me for each of the following statements whether you think it can always be justified, never be justified, or something in between: “Cheat- ing on tax if you have the chance”. García et al. (2018, p. 8) have identified that previous studies included two groups of research related to the EVS’s state- ment on cheating on taxes. The first group included research where higher scores involved higher intrinsic willingness to pay taxes (Torgler, 2003, p. 286; Torgler, 2012, p. 23; Lago-Peñas and Lago-Peñas, 2010, p. 441). In the second group authors examined ‘unwillingness to cheat’ (Alm et al., 2006, p. 861; Frey and Torgler, 2007, pp. 140–142). Academic research also acknowledges the measurement of individual tax morale through various survey questions. Ac- cording to social psychology theory (Lewis, 1982, p. 144), when individuals ob- serve widespread tax evasion or the behavior of their peers, their willingness to pay taxes may decline, diminishing intrinsic motivation and encouraging Central European Public Administration Review, Vol. 23, No. 1/2025 135 Examining Individual Tax Morale in Europe with Machine-Learning Methods opportunistic behavior (Torgler, 2005, p. 136). Hence, authors such as Torgler (2005, p. 136), Sá et al. (2013, p. 4) and Doerrenberg et al. (2014, p. 39) have analysed responses to the level of tax morale by including the statement of ‘avoidance of the payment of public transport fare’, or ‘manage to avoid paying all his tax’ in the Latinobarómetro (Torgler, 2005, p. 136) together with cheat- ing on tax statement. The research therefore involves two sets of modelling organized around two EVS statements: cheating on tax and avoidance of the payment of public trans- port fare. Since the full list of all articles where EVS has been previously aca- demically used is available on GESIS Leibniz Institute for the Social Sciences (Gesis, 2023), Appendix 1 provides an overview of the relevant findings. Ap- pendix 1 indicates measurement of individual tax morale through cheating on taxes statement (Tax morale represented as a dependent variable in Ap- pendix 1), its determinants, methods used and obtained results. Appendix 1 clearly indicates statistical and econometric methods and models that have exhaustively dealt with causality detection and identification of key factors of predicting individual tax morale in different circumstances – individual coun- tries, cross-country analyses, etc. Appendix 1 also indicates that determinants of individual tax morale are multifaceted, encompassing trust in institutions, perceived fairness, social capital, cultural background, socio-demographic fac- tors, economic context, and psychological considerations. Horodnic (2018, p. 870) provides systematic review of tax morale arguing that all factors could be grouped into three categories: formal institutions, informal institutions and socio-demographic characteristics and personal values. Under such cir- cumstances, tax morale is argued to be a result of the interaction between the formal and informal institutions. The possibility of ML’s application in social sciences and public policies has been recognized by Athey (2018, p. 509) and Lee (2020, p. 14), but not suf- ficiently applied. As the results from both Web of Science and Scopus topic research by two keywords of tax morale and machine learning generate only one result that examine the effects of religiosity and religion on individual tax morale (Davidescu et al., 2022, p. 2), ML application in social sciences is scarce. Most recently, Weber et al. (2018, p. 577) used ANN in the application regard- ing voter turnout and Chen et al. (2022, p. 1) used ANN together with Ordinal Logistic Regression method in predicting happiness levels of European im- migrants and natives. With further research, it has been determined that ML methods and mod- els have demonstrated substantial potential in enhancing various tax-related processes, but not prediction of individual tax morale. Most recent academic research indicates that by leveraging advanced algorithms and feature trans- formation techniques, ML models can significantly improve the accuracy of tax default predictions (Abedin et al., 2020, p. 19879), tax burden forecasts of agricultural enterprises (Kharitonova, 2023, p. 28), the detection of tax avoidance (Rahman et al., 2019, p. 536; Rahman et al., 2020, p. 722) and fraud (Murorunkwere et al., 2023, p. 731). These advancements not only help tax Central European Public Administration Review, Vol. 23, No. 1/2025 136 Lejla Lazović Pita, Almasa Odžak, Lamija Šćeta administrators and policymakers but also contribute to more efficient and informed financial decision-making. However, the ethical implications and po- tential biases of ML applications in these fields warrant careful consideration and ongoing research. Based upon current literature, a combination of formal and informal institu- tions together with socio-demographic and personal characteristics to pre- dict individual tax morale has been used by deploying ML methods. The mo- tive for this investigation arose from the fact that respondents across Europe in the EVS usually state that they would never cheat on taxes, but indicators in the economy such as tax evasion (e.g. VAT gap report) or corruption index tell a different story. Bearing in mind the significance of corruption as a vertical trust or a formal institution parameter (Horodnic, 2018, p. 873) or mechanism of reciprocity in tax morale as defined by Luttmer and Singhal (2014, p. 157), a further investigation on the significance of perceived levels of corruption on predicting individual tax morale has been conducted. Corruption in relation to tax morale has been somewhat investigated but mostly in Latin America, whereby Torgler (2005, p. 153) highlights that tax morale in Latin America is significantly affected by the shadow economy, corruption, and trust in of- ficials. Gerstenblüth et al. (2009, p. 2) and Gerstenblüth et al. (2012, pp. 129– 131) discuss how GDP per capita, income distribution, and corruption percep- tion shape tax morale in Latin America and the Caribbean while Jahnke and Weisser (2019, pp. 5–7) analyse how does petty corruption affect tax morale in Sub-Saharan Africa. Alm et al. (2016, p. 147) find that corruption among tax officials leads to higher levels of tax evasion of firms, with bribes reducing reported sales for taxes. Cung (2019, p. 189) shows that in Vietnam, economic freedom and corruption perceptions positively impact corporate income tax revenue, while inflation negatively impacts it. Most recently, Hsu (2023, p. 4) finds that when individuals perceive more government corruption, fiscal transparency is associated with lower tax morale. However, no research has yet integrated corruption perception index in assessing individual tax morale across Europe, as most European countries are among those with the lowest levels of perceived corruption worldwide. 3 Methods and Research Design 3.1 Machine Learning Methods and Models in The Individual Tax Morale Context Academic research across various disciplines suggests that the machine learn- ing (ML) methods and models presented in this subsection have a broad ap- plicability. After providing an overview of the six universal ML methods and models, their role in assessing individual tax morale will be contextualized and explored in the following subsections. The explanation will include the neces- sary research steps required, which will be presented in the results section. In the literature review section, the identified problem of predicting tax mo- rale has been investigated through regression analysis (measurement on a Central European Public Administration Review, Vol. 23, No. 1/2025 137 Examining Individual Tax Morale in Europe with Machine-Learning Methods scale with values from a continuous set, Appendix 1) or classification analy- sis (respondents are grouped into categories, Appendix 1). The article con- tributes to the latter. In terms of assessment of the level of individual tax morale, two target variables were selected: ‘Cheating on taxes’ (CoT) and ‘Avoiding fare of public transport’ (AFoPT). Respondents are classified into three groups based on their answers which reflect the degree of their agree- ment to those statements: ‘Never justifiable’, ‘Somewhat justifiable’, ‘Almost always justifiable’. So, a multiclass classification problem is defined. Thus, the aim is to predict the respondent’s group using available formal, informal, and socio-demographic data, framing the problem as a multiclass classification task in six ML methods and models. For the ML methods, contrary to binary classification problems, multiclass classifiers can distinguish between more than two classes. In order to solve the classification problem, a wide range of classical ML algorithms and meth- ods that are available in Scikit-Learn, an open-source ML library for Python, such as Logistic Regression, Random forests, k-Nearest Neighbors, etc. Also, deep learning frameworks in Python are used to build artificial neural net- work model using Keras and TensorFlow. It is important to note that before briefly explaining methods and models, some of the classical binomial classifiers can be easily adapted to multiclass classifiers (such as Decision trees and Random forests), while some others are based on multiple applications of binary classifiers when solving multiclass classification problems (such as logistic regression and Support vector clas- sifiers). When binary classification is used to solve multiclass classification problem, two strategies are available (Géron, 2022): the first strategy is called one-versus-the-rest (OvR) which uses one binary classifier for each class in the dataset, and the second strategy is called one-versus-one (OvO) which uses a binary classifier per each pair of classes in the dataset. Also, it is possible to use the cross-entropy loss if the ‘multi_class’ option is set to ‘multinomial’. Under Scikit-Learn, different strategies can be selected depending on applied algorithm (Yu et al., 2011). For a multiclass classification problem, six types of models were created and tested: the logistic regression (LR), the decision tree (DT), random forest (RF), k-Nearest Neighbors (kNN), Support vector classifier (SVC) and an artificial neural network (ANN) model providing explanation for each selected model. The logistic regression (LR) model is based on the estimation of the prob - ability that an instance belongs to a particular class using a logistic function. When using logistic regression for the multiclass classification problem dif- ferent solvers are available. A ‘newton-cg’ as a solver has been used, based on Newtons method, and parameter ‘multi_class’ is defined as ‘multinomial’ (Yu et al., 2011). The decision tree (DT) model is based on a tree (constructed from nodes and edges), as a hierarchical data structure. Each node in the constructed tree rep- resents a decision based on a feature, while each leaf node represents a class. Central European Public Administration Review, Vol. 23, No. 1/2025 138 Lejla Lazović Pita, Almasa Odžak, Lamija Šćeta The prediction is made by traversing through the model tree from the root to the leaf using input data. Random forest (RF) is a typical example of an ensemble model, i.e., it uses sev- eral models and aggregates their result. Precisely, the RF model uses many decision trees on random subsets of the features of the problem, then pro- duces a result by averaging out their predictions. k-Nearest Neighbors (kNN) model for classification estimates the class that an instance belongs to, using majority voting among the k-nearest neighbors of the given instance. The model assumes that the distance between the given instance and all the data in the training set is precalculated and that the hy- perparameter k is selected, thus k-nearest neighbors can be identified. Support vector classifier (SVC) is a model for solving multiclass classification problems based on maximizing the margin between the decision boundary and the closest samples from the different classes in the feature space of the data set. It applies OvO strategy by default. An artificial neural network (ANN) is a machine learning model inspired by the networks of biological neurons, hence it is based on the neural network. The basic idea of ANN is to learn complex relationships between features and outputs by simulating the behaviour of neurons in the human brain. Neural network consists of layers of adequately connected nodes (neurons) where the initial layer corresponds to the features, while the final layer con- tains information about the classes of the problem. Internal layers are usu- ally called hidden layers. Each neuron in a hidden layer receives inputs from the previous layer, performs some computations, and passes outputs to the next layer (Géron, 2022). The research in social sciences applies ANN for im- proving prediction accuracy (Schmidhuber, 2015; Weber et al., 2018, p. 580; Chen et al., 2022, p. 2). 3.2 Research Design and Data The statement about cheating on taxes if one would have the chance meas- ured on a scale from 1 to 10 within the EVS has been exhaustively applied as a measurement of individual tax morale across European countries (Appendix 1). The statement regarding avoidance of the payment of public transport fare has also been previously used in the academic research in an additional measurement of the level of individual tax morale (Torgler, 2005, p. 136; Sá et al., 2013, p. 4; Doerrenberg et al., 2014, p. 39). However, the statement about the avoidance of paying a public transport fare might be seen as either a representation of ‘civic attitudes’ or might even represent dissatisfaction of transport service (Doerrenberg et al., 2014, p. 40) with a possibility of paying a small penalty. The cheating on taxes statement in its formulation might be used to predict that a decline in tax attitudes or tax morale will lead to an increase in tax evasion (Lewis, 1982, p. 177; Torgler, 2005, p. 135), indicat- ing a potentially stronger representation of a decline in tax morale than the avoidance of public transport fare statement. Both statements have previ- Central European Public Administration Review, Vol. 23, No. 1/2025 139 Examining Individual Tax Morale in Europe with Machine-Learning Methods ously been used in the assessment and measurement of individual tax morale. The academic research shows that determinants that explain the level of tax morale across countries are groups of socio-demographic variables together with groups of subjective economic, institutional, and social variables (Lago- Peñas and Lago-Peñas, 2010, p. 441; Hofmann et al., 2017, p. 65). Alternative- ly, factors that influence tax morale may be grouped as formal institutions, informal institutions and socio-demographic factors (Horodnic, 2018, p. 870) whereby tax morale is a result of interaction between formal and informal institutions, highlighting the significance of reciprocity as defined by Luttmer and Singhal (2014, p. 157). Based upon literature review and the absence of the application of ML meth- ods and models in predicting individual tax morale, following research ques- tions have been defined: 1. Can individual tax morale across Europe be predicted more accurately by using ML methods and models? 2. Does corruption perception index contribute towards a more consistent and accurate ML predictions? 3. What are data-driven key determinants in predicting individual tax morale across Europe by using ML methods? The data for the analysis is compiled from the fifth wave EVS dataset of June 2022 for all participating European countries until that point in time (EVS, 2020). EVS was conducted from September 2017 until October 2021 and alike previous EVS waves included statements that are used in the measure- ment of the level of tax morale. The entire fifth wave EVS methodology cor- responds to EVS (2020) and includes: sampling methods of respondents older than 18 years, explanation of country-by-country data collection and inclusion of 35 participating countries (excluding Latvia which was added in 2021). Par- ticipating countries and sample size per country are provided in Appendix 2. The data-driven analysis of the two selected statements as target variables, should contribute towards the group of research where higher scores involve higher intrinsic willingness to pay tax (Torgler, 2003, p. 286; Torgler, 2012, pp. 23–24; Alm et al., 2006, pp. 858–860; Lago-Peñas and Lago-Peñas, 2010, p. 441; Garcia et al., 2018). Both statements were recorded on a ten-point scale (1= never justified, 10=always justified) and similar to previous academic research were grouped into a three-point scale: K 0 – Never justifiable (corresponds to 1 in the original scale), K 1 – Somewhat justifiable (corresponds to values 2–7 in the original scale), and K 2 – Almost always justifiable (corresponds to values 8-10 in the original scale). Even though ML methods can overcome unbalanced datasets, similar procedure of grouping variables into three categories had been applied in previous studies, such as Alm et al. (2006, p. 852) or Ryšavá and Zídková (2021, p. 391). The instances where the targeted variables were missing have been dropped from the dataset (1342 instances or 2.3% of the entire dataset). The dataset included EVS individual data of 56761 instances collected across 35 European countries from 2017 to 2020. Central European Public Administration Review, Vol. 23, No. 1/2025 140 Lejla Lazović Pita, Almasa Odžak, Lamija Šćeta Based upon literature and identified variables summarized in Appendix 1, description of selected variables (including socio-demographic, formal and informal institutions or vertical and horizontal trust variables respectively, Horodnic, 2018, pp. 872–876) are presented in Table 1. In case of missing val- ues, Multivariate Imputation by Chained Equation method as described in the pre-processing phase has been applied. Reciprocity mechanism of tax morale and measures for changing formal insti- tutions are necessary for building vertical trust in public authorities (Horodnic, 2018, p. 878). In line with research questions, the research includes Transparen- cy International’s Corruption Perception Index (CPI) as the most distinguished global index of corruption and a measure of a vertical trust. CPI indicates the perceived level of public sector corruption in a country, and it ranks countries on the scale from 0 to 100 (Transparency International, 2022). The lower the perceived level of corruption, the higher the CPI and vice versa. The created CPI ranks countries by using several macroeconomic variables. Sources of data and methodology are unrelated and different from the opinion statements from the EVS, so the CPI can be integrated in such research. CPI per each coun- try was collected for the same year as the year of EVS survey. As noted by Horodnic (2018, p. 873), wider corruption is associated with lower tax morale. The purpose and the motivation to include external country-level determi- nant such as CPI in the research rose from the fact that in EVS responses, individuals mostly replied that they would never cheat on taxes and would always pay a public transport fare which usually does not correspond to the estimated level of corruption within a country (individuals tend to underes- timate that they would cheat on taxes). This is the case even for European countries which traditionally have lowest levels of CPI. Previous academic re- search determining the relationship between EVS’s tax morale and perceived corruption with the application of statistical methods and models has been scarce and already discussed in the literature review section. 3.3 Pre-Processing Phase In the pre-processing phase some classical steps have been performed, in- cluding data cleaning, feature engineering, some scaling, and encoding cat- egorical variables. In the feature engineering step, some of the selected vari- ables are modified and aggregated as described in Table 1. Missing values are treated very carefully in the cleaning data process (Géron, 2022). Multivariate Imputation by Chained Equation (MICE) method is used to fill in the missing data. It is a statistical method based on multiple imputations used to produce final input values aiming to preserve relationships of variables in the original dataset and to reduce the amount of the bias. Scaling of some variables and one-hot encoding for categorical variables has been performed, as required for the application of ML algorithms. The final data set consisting of 56761 instances is split randomly into training and test dataset by 60%–40% respec- tively. The results are similar without significant deviations even if the data is split differently. Central European Public Administration Review, Vol. 23, No. 1/2025 141 Examining Individual Tax Morale in Europe with Machine-Learning Methods Table 1: Description of selected variables and coding of selected variables Variable Code in EVS Modified measurement units Country level Country S003 Nominal measure – Numeric code for Country Year of Survey S020 Scale measure 2017–2020 Individual-level Gender X001 0 – male, 1 – female Age X003 0 – original scale 18–29 1 – original scale 30 and more Education X025_01 0 – low education level corresponds to original scale 0, 1, 2, 3 1 – medium level education (original scale 4, 5, 6) 2 – high level education (original scale 7, 8) Marital status X007 1 – married (original scale 1) 0 – not married (original scale 2–8) Religiosity F028 1 – once a week or more (original scale 1, 2) 0 – less than once a week (original scale 3–8) Public vs. Private sector employment X052 1 – public institution (original scale 1) 0 – not public institution (original scale 2, 3, 4) No. of children X011 0 – no child (original scale 0) 1 – one or more children (original scale 1–5) Employment status X028 2 – employed (original scale 1, 2) 1 – self-employed (original scale 3) 0 – other (original scale 4–10) Confidence in Government E069_11 Scale 1 to 4: 1 – A great deal, 4 – None at all Confidence in Parliament E069_07 Scale 1 to 4: 1 – A great deal, 4 – None at all Confidence in Justice System/Courts E069_17 Scale 1 to 4: 1 – A great deal, 4 – None at all Bribery F117 Scale 1 to 10: 1 – Never justifiable, 10 – Always justifiable Government responsibility E037 Scale 1 to 10: 1 – Individuals should take more responsibility for providing for themselves, 10 – The state should take more responsibility to ensure that everyone is provided for Immigrants strain G041 Scale 1 to 10: 1 – Immigrants are a strain on a country’s welfare system, 10 – Immigrants are not a strain on a country’s welfare system Source: authors. Central European Public Administration Review, Vol. 23, No. 1/2025 142 Lejla Lazović Pita, Almasa Odžak, Lamija Šćeta 4 Results Separate analyses have been performed for each of the two target variables as measures of individual tax morale. They have been analysed through the two scenarios named “Country” and “CPI”. Initially, in the Country scenario, selected variables have been used, while in CPI scenario the ‘country’ variable has been replaced with the ‘CPI’. Having in mind that the level of tax morale measured by the two target vari- ables might not be free from biases and problems (Torgler and Valev, 2010), in the conducted ML analysis, it is necessary to include the CPI in all models as an external variable not accounted for in the EVS. The reason lies in the exami- nation if the level of prediction of individual tax morale would improve. Since cross-cultural comparisons should be used cautiously, especially in countries with high corruption and tax evasion (Torgler, 2005), the inclusion of the CPI can be justified from the speeding up reasons (Torgler and Valev, 2010). The six ML models and methods are applied for each target variable and for each scenario. Created models are evaluated and compared using accuracy, a commonly used metric to evaluate multiclass classification models. Results represented as accuracies are calculated for the training and test sets and are presented in Table 2. Table 2: ML methods performance of “Country” and “CPI” scenarios for target variables ‘CoT’ and ‘AFoPT’ Accuracy - CoT Accuracy - AFoPT Training set Test set Training set Test set Scenario Country CPI Country CPI Country CPI Country CPI LR 0.72205 0.71036 0.71733 0.70641 0.64925 0.62878 0.65118 0.62532 DT 0.99471 0.71036 0.63637 0.70641 0.99413 0.99325 0.56071 0.56195 RF(10) 0.97175 0.97158 0.71975 0.71790 0.97199 0.97146 0.63400 0.63083 RF(100) 0.99471 0.99401 0.73552 0.73548 0.99410 0.99324 0.65268 0.64506 kNN(10) 0.71385 0.71036 0.66426 0.70641 0.67486 0.65953 0.61097 0.58291 kNN(25) 0.69544 0.68449 0.67236 0.66294 0.64852 0.63084 0.61845 0.59092 kNN(50) 0.68549 0.67253 0.67373 0.66118 0.63701 0.61863 0.61709 0.59560 SVC 0.71192 0.70487 0.70786 0.70099 0.63205 0.62333 0.63096 0.61973 ANN 0.74131 0.74656 0.73056 0.74104 0.70575 0.65416 0.65684 0.64643 Source: authors. To understand and evaluate each of the universal six ML models and meth- ods, it is necessary to contextualize and explain them in the specific assess- Central European Public Administration Review, Vol. 23, No. 1/2025 143 Examining Individual Tax Morale in Europe with Machine-Learning Methods ment of the individual tax morale. In the two models – RF and kNN, the values in brackets next to RF and kNN in Table 2 indicate several hyperparameters scenarios. For RF model hyperparameter represents number of trees used to create model, while for the kNN hyperparameter k represents number of neighbors of a single instance used for the prediction. For the ANN models underlying network is created using a classical “half-rule” approach. Accord- ing to this rule, the number of neurons in each successive hidden layer is ap- proximately half the size of the previous layer. Thus, for the Country scenario, 3 hidden layers with 26, 13 and 7 neurons were used. For CPI scenario, num- ber of initial variables is considerably smaller since categorical variables for the variable ‘country’ are replaced with numerical variable CPI, thus hidden layer consists of 9 and 5 neurons. In both cases half of the instances in the test sets are used as a validation set. A significant difference in the accuracy in train and test sets for DT and RF models indicates that overfitting occurs, i.e., the model is too complex, and it captures the noise in the training data instead of the underlying patterns. Thus, other models are preferred or - in ML methodology – some pruning, i.e., limiting the depth of the tree, needs to be done. So, the two models were excluded from further analysis. Six ML methods and models were further deployed to see if the level of pre- diction of the tax morale across Europe may be improved by including the country-level CPI as one of the most important vertical trust factors that shape the tax morale (Horodnic, 2018, p. 878). As the research developed in terms of interlinking vertical trust factors with the individual tax morale, a further investigation from the CPI scenario was researched by separating instances into two groups based on the country level CPI, namely the High corruption (CPI less or equal to 50) and the Moderate corruption group (CPI equal or greater than 51). The complete procedure was repeated for the two groups separately. The data set for High corruption group consisted of 21974 instances, while Moderate corruption group consisted of 34757 instances and the results are presented in Table 3. As the results in Table 2 indicate, accuracies for all six ML methods are consist- ent for both variables of interest. Very small differences between accuracies on the test and training datasets indicate that overfitting is not present, and that all conducted ML models are adequate. Central European Public Administration Review, Vol. 23, No. 1/2025 144 Lejla Lazović Pita, Almasa Odžak, Lamija Šćeta Table 3: Models’ performance for High corruption and Moderate corruption groups for target variable ‘CoT’ and for target variable ‘AFoPT’ Accuracy ‘CoT’ Accuracy ‘AFoPT’ Training set Test set Training set Test set Group High corruption Moderate corruption High corruption Moderate corruption High corruption Moderate corruption High corruption Moderate corruption LR 0.69804 0.72303 0.68817 0.72203 0.59815 0.64670 0.60262 0.63917 kNN(10) 0.69342 0.71124 0.63823 0.67273 0.63721 0.67368 0.56473 0.59871 kNN(25) 0.66512 0.69461 0.64585 0.67632 0.61233 0.64622 0.57281 0.60374 kNN(50) 0.65276 0.68101 0.64152 0.67826 0.60133 0.63362 0.57952 0.60884 SVC 0.69023 0.71431 0.68191 0.71276 0.59603 0.64215 0.59044 0.63356 ANN 0.73128 0.76181 0.71741 0.75321 0.64886 0.68259 0.61086 0.66154 Source: authors. 5 Discussion After applying several ML methods and models, the results and a response to the first research question indicate and confirm (Schmidhuber, 2015) that ML methods and models may be used in the assessment of the individual tax morale. Among all six ML methods and models, ANN model provides superior results compared to all other ML methods and models in predicting individual tax morale (Table 2 and Table 3). The perception of corruption has been previously identified as one of the key factors affecting individual tax morale (e.g. Torgler, 2005, p. 139; OECD, 2019, p. 28). To assess the significance of CPI as a measurement of corruption in the data-driven evaluation of individual tax morale, another set of ML methods and models were conducted as an answer to the second research question as presented in Table 3. Previous empirical research indicated a close relation- ship between the level of corruption measured by the CPI (Torgler, 2005, p. 153; Torgler, 2006; Torgler and Valev, 2010; Garcia et al., 2018) or inclusion of the perceived bribery in the explanation of the two target variables. Bribery is a concept closely related to corruption (OECD, 2013) and in tax terminology related to tax evasion and tax audits, it is also considered a type of corruption. Furthermore, Vargas-Hernández (2009, p. 272) states that some examples of forms of corruption are bribery, collusion, embezzlement of public funds and theft, fraud, extortion, abuse of discretion, favouritism, clientelism, nepotism, the sale of government property by public officials, patronage, etc. The results from Table 3 reveal several important findings. Firstly, the results are consistent for both target variables across applied ML models which can Central European Public Administration Review, Vol. 23, No. 1/2025 145 Examining Individual Tax Morale in Europe with Machine-Learning Methods be verified by the best fitting results obtained in the ANN models. Secondly, if the results within training set or within test set for both target variables are compared, the results show an improvement of accuracy of predictions in the Moderate corruption group compared to High corruption group. Thirdly, the results remain consistently improved in the test set with reduced variability in comparison to the training set. Finally, and most importantly, the results for both target variables regardless of training and test set indicate an increase in accuracy of prediction in the Moderate corruption group where country-level CPI is higher, hence the level of perceived corruption is lower. The results are in line with previous literature as stated in Horodnic (2018, p. 878) and the response to the second research question is positive. In the broader context of the virtuous circle between effective government performance and tax mo- rale, the application of the six ML methods and models with the inclusion of the CPI as a measurement of corruption has several important policy impli- cations in the assessment of individual tax morale. As Torgler (2005, p. 135) and Sá et al. (2013, p. 1) found a negative correlation between tax morale and tax evasion and shadow economy, corruption has been recognized as a factor that affects the taxpayers in terms of lowering their tax morale (hence increas- ing tax evasion, Torgler, 2005, p. 139). Furthermore, since corruption also af- fects morality among the tax administration officials (Torgler, 2005, p. 139), if corruption occurs within public administration, it inevitably affects the entire tax system and therefore credibility of formal institutions (Horodnic, 2018, p. 871). It affects both efficiency of allocation in terms of possible delays in trans- actions and affects the fairness of the entire tax system, which in turn confirms the social psychology milestones set by Lewis (1982, p. 177) in diminishing in- trinsic motivation to pay taxes and encouraging opportunistic behaviour if the avoidance occurs and therefore might lead to evasion. The presented research investigated both – the intention to avoid paying a public transport fee as a proxy to avoidance issues, and the intention to evade taxes measured through a proxy of cheating on taxes statement and the obtained results confirmed that ML methods and models may be used for such predictions. If the obtained results are compared to the previous academic analyses of the differences between old and new EU member states in terms of the level of tax evasion (Yamen et al., 2018, p. 26; Dobrovič, Rajnoha, and Šuleř, 2021, p. 705), the data-driven results from six ML methods and models still confirm the existing differences in the level of individual tax morale among countries in their significance of a country belonging to a high versus a moderate CPI group. The results support the fact that in countries with higher CPI and therefore lower perceived corruption, the perceived level of individual tax morale can be predicted with higher accuracy, indicating higher respondent’s honesty when asked to estimate individual level of tax morale. The opposite is valid for responses in the High corruption group of countries with lower CPI and higher perceived corruption. These results further support the significance of reciprocity mechanism of tax morale (Luttmer and Singhal, 2014, p. 157). Similar conclusions, using different methods have been previously elaborated Central European Public Administration Review, Vol. 23, No. 1/2025 146 Lejla Lazović Pita, Almasa Odžak, Lamija Šćeta in Torgler (2006), Lago-Peñas and Lago-Peñas (2010), Garcia et al. (2018) and Horodnic (2018, p. 878). As presented in Table 3, the results indicate higher accuracy of prediction for the ‘CoT’ target variable than ‘AFoPT’ target variable across all models. This indicates superiority of the ‘CoT’ variable to ‘AFoPT’ variable in estimating and measuring individual tax morale and/or individual tax morale. A further expla- nation might be that ‘AFoPT’ is considered a less serious aspect of tax evasion due to its relatively small penalty and the fact that the sanction for non-com- pliance is only financial (payment of a fine/penalty). A possible limitation of the AFoPT statement might be in a fact that respondents indicated a dissat- isfaction with a transport service (Doerrenberg et al., 2014, p. 40). Therefore, the results by using ML methods and models are more accurate with the ‘CoT’ variable, that is considered a more fraudulent, with higher financial implica- tions that might also have criminal implications. Both of results are in line with previous studies such as James et al. (2019). To answer the final, third research question, the importance of the input variables for ANN models has been further analysed. The permutation impor- tance method is performed to calculate the feature importance of construct- ed models for the given dataset. Feature importance as mostly applied in ANN has been obtained in a such way that it can be understood as a decrease in a score associated with the model when values for the selected feature are shuffled randomly. The effects of random shuffling values are measured using an appropriate scoring function and the ‘Explained variance’ metric as a scoring function has been used. Even though the list may be more exhaustive, as a response to the third research question, the results indicate three most important determinants of individual tax morale: 1. The respondents’ attitudes towards accepting a bribe during their duties emerged as the most significant variable across all CPI scenarios and gro- ups within; 2. CPI was consistently the second most crucial variable, except in the case of CPI scenario (CoT) and Moderate corruption group, where it was replaced by the gender variable; 3. Gender together with respondents’ belief that immigrants impose a strain on a country’s welfare system were a third most important determinant of individual tax morale. On one side, these findings support the significance of interaction between the formal and informal institutions together with socio-demographic factors as indicated in the works of Horodnic (2018, p. 879). On the other, the results highlight the significance of reciprocity of tax morale (Luttmer and Singhal, 2014, pp. 157-160). Furthermore, in the European context, the results con- firm the virtuous circle between effective government performance, high tax morale and voluntary tax compliance (OECD, 2018) which are very important to the European tax administration officials and tax policy makers. Finally, the results contribute towards academic literature in the understanding of Central European Public Administration Review, Vol. 23, No. 1/2025 147 Examining Individual Tax Morale in Europe with Machine-Learning Methods human and economic behaviour in the investigation of individual tax morale across Europe by applying ML methods and models. Recommendations for future research include the application of other pos- sible variables in the present research, such as country’s belonging to the Eu- ropean Union (or not) together with other official macroeconomic country indicators. Obtained results by using different sources of data, for example Special Eurobarometer 498 – Wave EB92.1., World Values Survey, Afroba- rometer, Latinobarómetro, may also be tested using ML methods and mod- els which ought to provide useful insights and perhaps highlight important regional differences. Further research might also use target variables on the original scale, since ML methods available in Python can handle imbalanced datasets. As a research limitation, it is important to note that the accuracy of predictions in social sciences is somewhat below the results that can be ob- tained with ML methods and models in natural sciences or engineering. 6 Conclusions The primary purpose of the interdisciplinary study was in the contribution towards academic literature in the understanding of human and econom- ic behaviour in the investigation of individual tax morale across Europe by applying data-driven ML methods and models. Research was based on the dataset that included 56761 individual data selected from 35 countries and compiled from the EVS over 2017-2020 period. The study investigated the relationship between ‘CoT’ (cheating on taxes) and ‘AFoPT’ (avoiding public transport fare), as two target variables and proxies to individual tax morale, together with factors explained by sixteen sociodemographic and subjective economic, institutional, and social variables. Three research questions have been defined with several important policy level conclusions for government and tax administration officials across European countries and beyond. The research highlights the role of ML methods and models in explaining individ- ual tax morale, corruption levels, and broader human behaviour. The findings demonstrate that ML methods and models enhance accuracy and predictive power, particularly when the CPI is incorporated into the analysis. Among all applied ML methods and models, ANN models yielded the highest prediction accuracy. Furthermore, the results indicate that CPI is a stronger predictor than country-specific variables for the two proxy variables of individual tax morale, reinforcing the importance of corruption levels in tax morale assess- ments. The results again highlight the significant policy implications for the European decision makers, regulators together with the EU institutions (tax officials, tax administration and beyond, Horodnic, 2018, p. 871; OECD 2018; 2019) in their determination to research and tackle tax evasion and increase tax morale across EU. Finally, findings of this study contribute significantly to social sciences by enhancing the understanding of human behaviour measured by the key fac- tors shaping individual tax morale across Europe. The results reveal that re- spondents’ attitudes toward bribery, the CPI, beliefs about immigrants’ im- Central European Public Administration Review, Vol. 23, No. 1/2025 148 Lejla Lazović Pita, Almasa Odžak, Lamija Šćeta pact on national welfare, and gender are the most important determinants of tax morale. These insights align with previous social science research (e.g., Lago-Peñas and Lago-Peñas, 2010; Garcia et al., 2018), highlighting the impor- tance of these variables in shaping tax compliance behaviour. Future research should continue to incorporate ML methods and models in the assessment of the key factors influencing perceived tax morale, as they offer a deeper understanding of the social, economic, and ethical dimensions influencing in- dividuals’ willingness to comply with tax obligations. Central European Public Administration Review, Vol. 23, No. 1/2025 149 Examining Individual Tax Morale in Europe with Machine-Learning Methods References Abedin, M. Z. et al. (2020). Tax default prediction using feature transformation- based machine learning. IEEE Access, 9, pp. 19864–19881. https://doi.org/10. 1109/ACCESS.2020.3048018 Allingham, M.G. and Sandmo, A. (1972). Income tax evasion: A theoretical analysis. Journal of public economics, 1(3-4), pp. 323–338. http://doi.org/10. 1016/0047-2727(72)90010-2 Alm, J. (2014). Expanding the theory of tax compliance from individual to group motivations. In F. Forte et al., eds., A Handbook of Alternative Theories of Public Economics. Edward Elgar Publishing, pp. 260–277. Alm, J. and Gomez, J. L. (2008). Social capital and tax morale in Spain. Economic Analysis and Policy, 38(1), pp. 73–87. Alm, J. and Malézieux, A. (2021). 40 years of tax evasion games: a meta-analysis. Experimental Economics, 24(3), pp. 699–750. https://doi.org/10.1007/s10683 -020-09679-3 Alm, J. and Torgler, B. (2006). Culture differences and tax morale in the United States and in Europe. Journal of economic psychology, 27(2), pp. 224–246. https://doi.org/10.1016/j.joep.2005.09.002 Alm, J., Martinez-Vazque, J. and Torgler, B. (2006). Russian attitudes toward paying taxes–before, during, and after the transition. International Journal of Social Economics 33(12), pp. 832–857. https://doi.org/10.1108/030682906 10714670 Alm, J., Martinez-Vazquez, J. and McClellan, C. (2016). Corruption and firm tax evasion. Journal of Economic Behavior & Organization, 124, pp. 146–163. https://doi.org/10.1016/j.jebo.2015.10.006 Alm, J., Sanchez, I. and De Juan, A. (1995). Economic and noneconomic factors in tax compliance. Kyklos, 48(1), pp. 3–18. https://doi.org/10.1111/j.1467-64 35.1995.tb02312.x Andreoni, J., Erard, B. and Feinstein, J. (1998). Tax compliance. Journal of economic literature, 36(2), pp. 818–860. Athey, S. (2018). The impact of machine learning on economics. In A. Agrawal, J. Gans and A. Goldfarb, eds., The economics of artificial intelligence: An agenda. University of Chicago Press, pp. 507–547. Barone, G. and Mocetti, S. (2011). Tax morale and public spending inefficiency. International Tax and Public Finance, 18, pp. 724–749. Becker, G.S. (1968). Crime and punishment: An economic approach. Journal of political economy, 76(2), pp. 169–217. Castañeda, V. M. (2021). Tax Equity and its Association with Fiscal Morale. International Public Management Journal, 24(5), pp. 710–735. Chen, S., Yang, M. and Lin, Y. (2022). Predicting happiness levels of European immigrants and natives: An application of Artificial Neural Network and Ordinal Logistic Regression. Frontiers in Psychology, 13. https://doi.org/10.33 89/fpsyg.2022.1012796 Cung, N. H. (2019). Impact of economic freedom index and corruption percep- tions index on corporate income tax revenue in Vietnam. European Scientific Journal, 15(28), pp. 185–196. https://doi.org/10.19044/esj.2019.v15n28p1 Cyan, M.R., Koumpias, A.M. and Martinez-Vazquez, J. (2016). The determinants of tax morale in Pakistan. Journal of Asian Economics, 47, pp. 23–34. Central European Public Administration Review, Vol. 23, No. 1/2025 150 Lejla Lazović Pita, Almasa Odžak, Lamija Šćeta Davidescu, A.A. et al. (2022). Could Religiosity and Religion Influence the Tax Morale of Individuals? An Empirical Analysis Based on Variable Selection Methods. Mathematics, 10(23), p. 44–97. https://doi.org/10.3390/math102 34497 Dobrovič, J., Rajnoha, R. and Šuleř, P. (2021). Tax evasion in the EU countries following a predictive analysis and a forecast model for Slovakia. Oeconomia Copernicana, 12(3), pp. 701–728. https://doi.org/10.24136/oc.2021.023 Doerrenberg, P. et al. (2014). Nice guys finish last: Do honest taxpayers face higher tax rates? Kyklos, 67(1), pp. 29–53. https://doi.org/10.1111/kykl.12042 Dulleck, U. et al. (2016). Tax compliance and psychic costs: Behavioral experimental evidence using a physiological marker. Journal of Public Economics, 134, pp. 9–18. EVS, 2020. European values study 2017: Integrated dataset (EVS 2017). GESIS Data Archive, Cologne. ZA7500 data file version 4.0.0. At , accessed May 2023. Feld, L.P. and Frey, B.S. (2002). Trust breeds trust: How taxpayers are treated. Economics of governance, 3, pp. 87–99. Frey, B.S. and Torgler, B. (2007). Tax morale and conditional cooperation. Journal of comparative economics, 35(1), pp. 136–159. https://doi.org/10.1016/j.jce. 2006.10.006 García, G.A., Azorín, J.D.B. and de la Vega, M.D.M.S. (2018). Tax evasion in Europe: An analysis based on spatial dependence. Social Science Quarterly, 99(1), pp. 7–23. https://doi.org/10.1111/ssqu.12382 Géron, A. (2022). Hands-on machine learning with Scikit-Learn, Keras, and TensorFlow. O’Reilly Media, Inc. Gerstenblüth, M. et al. (2009.) Threats in Latin American and Caribbean countries: how do inequality and the asymmetries of rules affect tax morale?. Documento de Trabajo/FCS-DE; 14/08. https://doi.org/10.2139/ssrn.1489393 Gerstenblüth, M et al. (2012). How do inequality affect tax morale in Latin America and Caribbean?. Revista de Economía del Rosario, 15(2), pp. 123–135. GESIS Leibniz Institute for the Social Sciences (2023). EVS-Bibliographs. GESIS Data Archive, Cologne. At , accessed 4 July 2024. Hofmann, E. et al. (2017). Tax compliance across sociodemographic categories: Meta-analyses of survey studies in 111 countries. Journal of Economic Psychology, 62, pp. 63–71. https://doi.org/10.1016/j.joep.2017.06.005 Horodnic, I. A. (2018). Tax morale and institutional theory: a systematic review. International journal of sociology and social policy, 38(9/10), pp. 868–886. Hsu, H. Y. (2023). Fiscal transparency and tax morale: is the relationship shaped by perceptions of government performance and corruption? International Review of Administrative Sciences, pp. 1–18. https://doi.org/10.1177/00208 523231220599 Jahnke, B. and Weisser, R. A. (2019). How does petty corruption affect tax morale in Sub-Saharan Africa? European Journal of Political Economy, 60, p. 101751. https://doi.org/10.1016/j.ejpoleco.2018.09.003 James, S. et al. (2019). How seriously do taxpayers regard tax evasion? A survey of opinion in England. Journal of Money Laundering Control 22(3), pp. 563– 575. Central European Public Administration Review, Vol. 23, No. 1/2025 151 Examining Individual Tax Morale in Europe with Machine-Learning Methods Kharitonova A.E. (2023). Forecasting the tax burden of agricultural enterprises by machine learning methods. Taxes and Taxation, 4, pp. 28–38. DOI: 10.7256/2454-065X.2023.4.43917 Kountouris, Y. and Remoundou, K. (2013). Is there a cultural component in tax morale? Evidence from immigrants in Europe. Journal of Economic Behavior and Organization, 96, pp. 104–119. Lago-Peñas, I. and Lago-Peñas, S. (2010). The determinants of tax morale in comparative perspective: Evidence from European countries. European Journal of Political Economy, 26(4), pp. 441–453. https://doi.org/10.1016/j. ejpoleco.2010.06.003 Lee, J.A. (2020). Artificial Neural Networks in Public Policy: Towards an Analytical Framework. Doctoral dissertation. George Mason University Lewis, A. (1982). The psychology of taxation. Oxford: Martin Robertson. Listhaug, O. and Miller, A.H. (1985). Public Support for Tax Evasion: Self-interest or Symbolic Politics? European Journal of Political Research, 13(3), pp. 265– 282. http://doi.org/10.1111/j.1475-6765.1985.tb00123.x Lubian, D. and Zarri, L. (2011). Happiness and tax morale: An empirical analysis. Journal of Economic Behavior & Organization, 80(1), pp. 223–243. Luttmer, E.F. and Singhal, M. (2014). Tax morale. Journal of economic perspectives, 28(4), pp. 149-168. Molero, J.C. and Pujol, F. (2012). Walking inside the potential tax evader’s mind: tax morale does matter. Journal of Business Ethics, 105, pp. 151–162. Murorunkwere, B. F. et al. (2023). Predicting tax fraud using supervised machine learning approach. African Journal of Science, Technology, Innovation and Development, 15(6), pp. 731–742. https://doi.org/10.1080/20421338.2023.2 187930 OECD (2013). Bribery and Corruption Awareness Handbook for Tax Examiners and Tax Auditors. OECD Publishing, Paris. https://doi.org/10.1787/97892642 05376-en. OECD (2019). Tax Morale: What Drives People and Businesses to Pay Tax? OECD Publishing, Paris. https://doi.org/10.1787/f3d8ea10-en OECD/CAF/ECLAC (2018). Latin American Economic Outlook 2018: Rethinking Institutions for Development. OECD Publishing, Paris. https://doi. org/10.1787/leo-2018-en. Rahman, R. A., Masrom, S. and Omar, N. (2019). Tax avoidance detection based on machine learning of Malaysian government-linked companies. International Journal of Recent Technology and Engineering, 8(2), pp. 535– 541. DOI: 10.35940/ijrte.B1083.0982S1119 Rahman, R. A. et al. (2020). An application of machine learning on corporate tax avoidance detection model. IAES International Journal of Artificial Intelligence, 9(4), pp. 721–725. DOI: 10.11591/ijai.v9.i4. Riahi-Belkaoui, A. (2004). Relationship between tax compliance internationally and selected determinants of tax morale. Journal of international accounting, auditing and taxation, 13(2), pp. 135–143. Richardson, G. (2006). Determinants of tax evasion: A cross-country investigation. Journal of international Accounting, Auditing and taxation, 15(2), pp. 150–169. Ryšavá, T. and Zídková, H. (2021). What are the factors of tax evasion? New findings in the EVS Study. Review of Economic Perspectives, 21(4), pp. 385– 409. https://doi.org/10.2478/revecp-2021-0017 Central European Public Administration Review, Vol. 23, No. 1/2025 152 Lejla Lazović Pita, Almasa Odžak, Lamija Šćeta Sá, C., Martins, A. and Gomes, C. (2013). Tax morale, occupation and income level: An analysis of portuguese taxpayers. Comunicações. Schmidhuber, J. (2015). Deep learning in neural networks: An overview. Neural Networks, 61, pp. 85–117. https://doi.org/10.1016/j.neunet.2014.09.003 Schmölders, G. (1970). Survey research in public finance: A behavioral approach to fiscal theory. Public finance, 25(2), pp. 300–306. Schneider, F., Raczkowski, K. and Mróz, B. (2015). Shadow economy and tax evasion in the EU. Journal of Money Laundering Control, 18(1), pp. 34–51. Strümpel, B. (1969). The contribution of survey research to public finance. In A. T. Peacock, ed., Quantitative Analysis in Public Finance. New York: Praeger Publishers, pp. 14–32. Torgler, B. (2003). To evade taxes or not to evade: that is the question. The Journal of Socio Economics, 32(3), pp. 283–302. https://doi.org/10.1016/ S1053-5357(03)00040-4 Torgler, B. (2005). Tax morale in Latin America. Public choice 122(1-2), pp. 133–157. Torgler, B. (2006). The importance of faith: Tax morale and religiosity. Journal of economic Behavior & organization, 61(1), pp. 81–109. https://doi. org/10.1016/j.jebo.2004.10.007 Torgler, B. (2012). Tax morale, eastern Europe and European enlargement. Communist and Post-Communist Studies, 45(1-2), pp. 11–25. https://www. jstor.org/stable/48609657 Torgler, B. and Schneider, F. (2005). Attitudes towards paying taxes in Austria: An empirical analysis. Empirica, 32(2), pp. 231–250. https://doi.org/10.1007/ s10663-004-8328-y Torgler, B. and Schneider, F. (2007). What shapes attitudes toward paying taxes? Evidence from multicultural European countries. Social Science Quarterly, 88(2), pp. 443–470. https://www.jstor.org/stable/42956304 Torgler, B. and Valev, N.T. (2010). Gender and public attitudes toward corruption and tax evasion. Contemporary Economic Policy, 28(4), pp. 554–568. https:// doi.org/10.1111/j.1465-7287.2009.00188.x Transparency International. (2022). Corruption Perception Index. Retrieved from transparency.org/en/cpi/2022 Vargas-Hernández, J.G. (2009). The multiple faces of corruption: Typology, forms and levels. In A. Stachowicz-Stanusch, ed., Organizational immunity to corruption: Building Theoretical and Research Foundations. Information Age Publishing Inc.: Charlotte, pp. 111–117. Weber, P. et al. (2018). Prospect for knowledge in survey data: an artificial neural network sensitivity analysis. Social Science Computer Review, 36(5), pp. 575– 590. https://doi.org/10.1177/0894439317725836 Yamen, A. et al. (2018). Impact of institutional environment quality on tax evasion: A comparative investigation of old versus new EU members. Journal of International Accounting, Auditing and Taxation, 32, pp. 17–29. https:// doi.org/10.1016/j.intaccaudtax.2018.07.001 Yitzhaki, S. (1974). A note on income tax evasion: A theoretical analysis. Journal of public economics, 3, pp. 201–202. http://doi.org/10.1016/0047- 2727(74)90037-1 Yu, H.F., Huang, F.L. and Lin, C.J. (2011). Dual coordinate descent methods for logistic regression and maximum entropy models. Machine Learning, 85, pp. 41–75. https://doi.org/10.1007/s10994-010-5221-8 Central European Public Administration Review, Vol. 23, No. 1/2025 153 Examining Individual Tax Morale in Europe with Machine-Learning Methods Appendix 1. Overview of the most relevant findings regarding measurement of individual tax morale Author(s) Methods/Data Dependent Variable Independent Variables Results Listhaug and Miller (1985) Regression analysis Tax morale Political ideology, religious values, political dissatisfaction, self-interest, personal dissatisfaction, political interest Strongest support for symbolic politics model; weak support for self-interest explanation Torgler (2005) Empirical study (Latinobarómetro and World Values Survey) Tax morale Size of shadow economy, tax burden, lack of honesty, corruption, trust in officials, belief in law enforcement, pro- democracy attitudes Tax morale is influenced by the size of the shadow economy, with tax burden, lack of honesty, and corruption being key factors; trust in officials, belief in law enforcement, and pro-democracy attitudes positively impact tax morale Torgler and Schneider (2005) Weighted probit estimation Tax morale Societal institutions such as trust or pride Decrease in tax morale between 1990 and 1999 in Austria; societal institutions are key determinants Richardson (2006) Cross-country investigation (OLS regression analysis) Tax evasion - TEVA - country survey rating of tax evasion collected by the World Economic Forum Complexity, education, income source, fairness, tax morale Lower complexity, higher education, income source, fairness, and tax morale are associated with lower tax evasion levels across 45 countries Torgler and Schneider (2007) Ordered probit model Tax morale Cultural and institutional differences Differences in tax morale exist between Switzerland, Belgium, and Spain Alm and Gómez (2008) Unique dataset (Spain’s Survey of Fiscal Policy) Tax morale Social capital, perception of benefits from public goods and services, perceived fiscal fraud Social capital and perception of benefits from public goods and services significantly influence tax morale in Spain, while perceived fiscal fraud negatively affects it Lago-Peñas and Lago- Peñas (2010) Multilevel model Tax morale Socio-demographic characteristics, personal financial experiences, political attitudes, regional GDP, tax arrangements, ethnic and linguistic fractionalizations Tax morale in European countries varies with socio-demographic characteristics, personal financial experiences, political attitudes, regional GDP, tax arrangements, ethnic and linguistic fractionalizations Torgler and Valev (2010) Weighted probit modelling Tax morale Gender, public attitudes towards corruption and tax evasion Women are less likely to justify corruption and tax evasion Central European Public Administration Review, Vol. 23, No. 1/2025 154 Lejla Lazović Pita, Almasa Odžak, Lamija Šćeta Lubian and Zarri (2011) Non-RCT observational study Stated happiness Fiscal honesty Tax morale, the intrinsic motivation to pay taxes, is a new determinant of happiness, as fiscal honesty generates a higher hedonic payoff than cheating Molero and Pujol (2012) Empirical study (binomial logit model) Tax morale Grievances about taxes, public funds, underground economic activities, duty, solidarity Tax morale is influenced by grievances about taxes, public funds, and underground economic activities, with a lesser impact from duty and solidarity Torgler (2012) Weighted ordered probit model Tax morale Socio-demographic, economic, social, political, and institutional variables Significant decrease in tax morale in 7 out of 10 Eastern European countries between 1999 and 2008 Kountouris and Remoundou (2013) Empirical study (European Social Survey) Tax morale Culture (immigrant origin country tax morale) Culture significantly influences individual tax morale, with immigrant origin country tax morale significantly influencing morale in the destination country Doerrenberg et al. (2014) Ordinary Least Squares (OLS) Tax morale Tax rates, avoiding public transport fee (instrumental variable) Higher tax morale groups bear a higher tax burden Cyan et al. (2016) Binary probit regression model Tax morale Education, gender, age, location (industrialized population centres) Tax morale in Pakistan is higher among educated individuals, especially those with very low or very high education, and is highest in major industrialized population centres; females generally show higher tax morale than males Horodnic (2018) Systematic review Tax morale Trust (vertical and horizontal), formal and informal institutions Horizontal and vertical trust are the most significant factors that positively influence tax morale García et al. (2018) Generalized linear model Tax morale Socio-demographic factors, subjective economic, social, institutional variables Impact of spatial dependence, economies of agglomeration, income inequality, economic imbalances, and perceived corruption on the variable -rejection of tax evasion Castañeda (2021) Pooled cross- sectional data Tax morale Horizontal equity, vertical equity Tax equity, including horizontal and vertical equity, is the most important determinant of fiscal morale Source. Own interpretation. Central European Public Administration Review, Vol. 23, No. 1/2025 155 Examining Individual Tax Morale in Europe with Machine-Learning Methods Appendix 2. Participating countries and sample size per country ISO Country code EVS No.of respondents ISO Country code EVS No.of respondents ISO Country code EVS No.of respondents ISO Country code EVS No.of respondents ISO Country code EVS No.of respondents ISO Country code EVS No.of respondents ISO Country code EVS No.of respondents DK 3362 DE 2170 GB- GBN 1788 UA 1612 RS 1499 PL 1352 SE 1194 CH 3174 FR 1870 BA 1724 BG 1558 HR 1487 EE 1304 NO 1122 NL 2404 RU 1825 AT 1644 BY 1548 LT 1448 PT 1215 MK 1117 IT 2277 CZ 1811 IS 1624 HU 1514 AL 1435 ES 1209 SI 1075 GE 2194 AZ 1800 RO 1613 AM 1500 SK 1432 FI 1199 ME 1003 Source: EVS, 2020.