Geração de dados sintéticos longitudinais a partir de estruturas causais

Abstract

Causal inference seeks to identify cause-and-effect relationships, going beyond corre-lation by estimating how outcomes would change under different conditions. This ability to predict counterfactual outcomes is fundamental in real-world applications — such as medicine, finance, and the social sciences — where reliable decisions depend on a causal understanding of phenomena. The systematic evaluation of causal inference models is still limited by the scarcity of benchmark datasets whose underlying mechanisms are fully known. In this context, this dissertation presents the Causal Synthetic Data Generator (CSDG), an open-source tool capable of generating synthetic longitudinal data governed by causal structures with explicit autoregressive dynamics. The CSDG enables detailed control over the strength of variable effects, treatment interventions, and noise levels, offering a flexible and domain-independent platform for the experimentation and evaluation of causal learning algorithms. The proposed forma-lization is based on autoregressive structural causal equations, integrating cause-effect relationships with temporal dependencies and allowing the generation of both factual and counterfactual scenarios under different causal structures. To validate the consistency of the generated data, this work includes a quantitative analysis stage and experiments on a prediction task. The evaluations show that the synthetic data preserve expected properties of correlation and response to interventions, and the results demonstrate both causal coherence and statistical realism, confirming the suitability of the CSDG as a benchmarking and validation tool for causal models. The CSDG source code is publicly available at <https://github.com/angeruzzi/causal-synthetic-data-gen> .