Abstract

The persistent decline in voter turnout observed across established democracies poses important challenges for both substantive political analysis and statistical modeling. Italy represents a relevant case, as the 2022 parliamentary election registered the lowest turnout in the country’s republican history. From a methodological standpoint, modeling electoral participation requires statistical frameworks capable of handling complex interdependencies among socio-economic, health-related, and political variables, often inadequately captured by standard regression-based approaches.
This paper analyzes the determinants of voter turnout in the 2022 Italian parliamentary election using individual-level data from the European Social Survey (ESS). A rich set of variables is considered, including socio-demographic characteristics, income and economic status, general health, subjective well-being, trust in institutions, political attitudes, and indicators of political engagement. The main contribution of the paper is methodological: it proposes and illustrates a graphical modeling strategy based on Bayesian networks (BNs) to complement and extend conventional logistic regression analysis.
The empirical analysis proceeds in two steps. First, a logistic regression model is estimated to assess the marginal effects of the covariates on the binary outcome indicating whether the respondent voted. Average marginal effects are used to facilitate interpretation and comparison with existing results in the turnout literature. The regression results confirm well-established associations between turnout and age, education, Internet use, and institutional trust. In addition, significant effects of general health and life satisfaction are identified, highlighting the role of individual well-being in electoral participation.
Second, the analysis moves beyond the additive structure of the logistic model by adopting a BN framework, which encodes conditional independence assumptions through a directed acyclic graph (DAG). The network structure is learned from the data using the PC algorithm, a constraint-based method grounded in conditional independence testing. Structural learning is carried out under logically and temporally motivated constraints reflecting the hierarchical ordering of profiling variables, well-being and socio-political attitudes, and voting behavior. This approach allows the joint distribution of all variables to be factorized according to the learned DAG.
The estimated BN highlights the central role of income status and general health as direct parents of voter turnout, while also revealing indirect pathways operating through political attitudes and party proximity. Model performance is evaluated using receiver operating characteristic (ROC) analysis, with area under the curve (AUC) values indicating excellent predictive accuracy.
An additional methodological advantage of BNs is their suitability for scenario-based inference. By conditioning on specific configurations of income and health status, the model quantifies their joint and non-additive effects on the probability of voting. The results show that the coexistence of economic hardship and poor health substantially amplifies the likelihood of abstention compared to scenarios in which only one disadvantage is present.
Overall, Bayesian networks provide a flexible and statistically rigorous tool for modeling voter turnout, enabling the explicit analysis of dependency structures, indirect effects, and joint influences that are not accessible through traditional regression models. The proposed approach has broader applicability for the analysis of complex social survey data in political and social statistics.