Mechanisms beyond synthetic worlds: foundations of empirical agent-based models

Contemporary societal phenomena--such as climate crisis, political radicalization, and polarization--require understanding human behavior from a collective perspective. Agent-based models (ABMs) provide mechanistic descriptions of the emergence of macroscopic phenomena from individual actions and interactions. Leveraging bottom-up frameworks, ABMs capture the effect of microscopic behaviors on collectivities. However, ABMs are typically constrained within synthetic environments, and their properties are studied through forward simulations. This work proposes novel methodologies to build empirical ABMs. Empirical ABMs move beyond simulated worlds by unifying domain-specific theory and observed data within a single mechanistic framework. On the one hand, domain-specific theory encodes the behavioral mechanisms under study, and it is usually grounded in social science, economics, biology, or other disciplines. On the other hand, observed data enable an empirical approach by providing a quantitative representation of the phenomena of interest. The construction of an empirical ABM is the process of developing a model that describes the domain-specific theory by representing the observed data. Such construction involves three main tasks: (i) mechanism identification, (ii) model initialization, and (iii) model fitting. First, mechanism identification unveils the mechanisms of the theoretical assumptions that are observable in the data. I present a causal model that integrates domain-specific theory to explain complex social dynamics in online social networks. By connecting the formalism of causal models and ABMs, I provide a framework to identify the cause-effect relationships that are both hypothesized by theory and observed in data. Second, model initialization involves the generation of synthetic data that resemble the data at the initial snapshot. In this thesis, I present a population synthesis with geographic coordinates and a novel evaluation pipeline. Third, model fitting estimates the model parameters that best describe the system evolution observed in the data. In this thesis, I review and synthesize the literature on parameter estimation in ABMs, unifying approaches from different disciplines under a common taxonomy. Beyond revisiting existing literature, I propose novel approaches to model fitting based on the formalism of Probabilistic Generative ABMs (PGABMs). By using modern tools from probabilistic machine learning, the proposed approaches improve simulation-based inference both from a theoretical and an experimental perspective. Specifically, I employ Maximum Likelihood Estimation for statistical rigor and Variational Inference as an efficient, flexible alternative for parameter estimation within the PGABM framework. The joint contribution of mechanism identification, model initialization, and model fitting paves the way for multidisciplinary applications of empirical ABMs, thus enabling novel approaches to scientific discovery. In doing so, this thesis advances an explanatory paradigm where empirical ABMs do not simply mimic outcomes, but rather formalize the causal mechanisms necessary for a true scientific understanding of collective phenomena.