Causal Inference Beyond Standard Assumptions: Learning Policies and Treatment Effects in Complex Environments

Abstract

Causal inference aims to uncover cause-and-effect relationships from data and has seen widespread application and rapid methodological development across scientific disciplines. While recent methodological advances --- driven by increasingly rich and diverse data ---- has expanded the scope of causal analysis, practical applications often involve complexities that violate the core assumptions underlying standard approaches. These include lack of overlap between treatment and control groups, interference among units, and distributional shifts across populations. Addressing these challenges is crucial for ensuring the validity and reliability of causal conclusions in real-world settings. This dissertation develops novel methodological frameworks for robust, efficient, and interpretable causal inference in complex environments. Each chapter addresses a unique violation of standard assumptions, with an overall focus on two key areas: policy learning and heterogeneous treatment effect estimation. Chapter 1 considers safe policy learning in regression discontinuity designs, where treatment assignment is deterministic and requires robust extrapolation beyond observed data. Chapter 2 focuses on the evaluation and learning of individualized treatment rules under clustered network interference, where spillover effects exist and may vary across units within a cluster. Chapter 3 investigates the generalization of heterogeneous treatment effects with multisite data, where distributional shifts across populations challenge the validity of pooled or site-specific estimators.