Avançando a estimação robusta e confiável de efeitos de tratamento heterogêneos: inovações metodológicas e avaliações críticas

Abstract

Context: The robust and reliable estimation of Heterogeneous Treatment Effects (HTEs) is crucial across many scientific disciplines, yet it faces significant methodological challenges including model complexity, confounding, computational burden, and rigorous evaluation practices. Objectives: This thesis aimed to advance the field of HTE estimation by developing and critically evaluating methodologies that enhance the rigor, efficiency, and practical utility of causal inference techniques. Methodological Contributions: The work presents three primary contributions: (1) An empirical validation of the importance of ablation studies for complex nonparametric causal models, specifically examining the Bayesian Causal Forest (BCF) and the role of its propensity score component; (2) The development of the Test-Informed Simulation Count Algorithm (TISCA), a principled approach for determining the necessary number of replications in simulation studies for model evaluation using statistical principles; and (3) The introduction of the Differencein-Differences Bayesian Causal Forest (DiD-BCF), a novel non-parametric estimator for robust causal inference in DiD settings, particularly effectively addressing treatment effect heterogeneity through a Parallel Trends Assumption (PTA)-based reparameterization. Principal Findings: Ablation studies revealed that the propensity score component in BCF is not essential for performance and its omission can reduce computation time by approximately 21%. TISCA was shown to provide statistically justified simulation counts, promoting efficiency and enhancing the credibility of comparative model evaluations. DiDBCF demonstrated considerably superior performance over established benchmarks and uncovering nuanced conditional treatment effects in an empirical application to U.S. minimum wage policy. Overall Conclusion and Implications: This thesis collectively champions a paradigm of increased rigor, efficiency, and nuanced understanding in HTE estimation. It provides researchers with critically evaluated insights and novel tools— ablation study advocacy, a statistically grounded simulation design algorithm, and an advanced non-parametric DiD estimator—to generate more robust, reliable, and actionable causal evidence, thereby strengthening the foundation for evidence-based decision-making across various disciplines.