UC Berkeley

This dissertation studies methodological challenges in modern causal inference with a particular focus on the trustworthiness and robustness of estimation methods. While recent advances in machine learning and statistical methodology have vastly expanded the scope and toolkit of causal analysis, many questions remain unresolved. In practice, researchers often make causal conclusions despite unobserved counterfactuals, possible unmeasured confounding, as well as covariate imbalance in observational studies. First, unlike supervised prediction problems, since causal estimands such as treatment effects are never directly observed, model choice and uncertainty quantification become substantially more difficult in the absence of ground truth against which competing methods can be compared. Second, the assumption of no unobserved confounding is implausible in practice and existing sensitivity analysis methods seek to quantify how violations of this assumption could affect causal conclusions. However, these methods are typically designed for specific estimation strategies, motivating the need for a more general framework. Third, existing randomization inference methods create a mismatch between the design stage and the inference stage of matched observational studies, and sensitivity analyses for such studies are often inadequate. In the face of these problems, the three chapters of this dissertation develop methodological tools addressing these difficulties, with the common goal of improving the stability, robustness, and trustworthiness of causal inference in modern data settings.Chapter 1 develops a framework for stable model screening and uncertainty quantification for conditional average treatment effect (CATE) estimation - PCS-CATE. This is a joint work with Sizhu Lu and Bin Yu. The chapter adapts the Predictability–Computability–Stability (PCS) framework for veridical data science of Yu and Kumbier to causal inference settings. Because standard prediction-based validation is unavailable in causal problems, we adopt subgroup calibration metrics and a multi-metric screening procedure for evaluating CATE estimators. We further introduce a bootstrap-based uncertainty quantification framework that propagates both algorithmic and data uncertainty while calibrating interval estimates through subgroup coverage criteria. The framework is extended from randomized experiments to observational studies through additional propensity-score model diagnostics and screening procedures. Extensive simulations on synthetic and semi-synthetic data demonstrate that the proposed PCS procedure achieves more stable model screening and reliable interval coverage with mild width inflation relative to existing approaches.Chapter 2 studies the connection between sensitivity analyses of regression and weighting estimators, and formulates a unified framework of examining the sensitivity to unobserved confounding based on the perturbation of implied weights. This chapter is a joint work with Sam Pimentel and Melody Huang. Existing sensitivity analysis methods are often tailored to specific estimation strategies, such as partial R2 for linear regressions, or variance-based sensitivity model for IPW estimators. With the implied weight form for regression estimators, we constrain the L2 distance between implied weights based on observable covariates and oracle weight vector and characterize the resulting worst-case bias. We introduce two complementary formulations, the restricted model, which leads to interpretable sensitivity parameters and sharp bias bounds within this structured class, and the unrestricted model, which provides bounds that are robust to model misspecification. The framework naturally extends to interacted regressions and other estimators that admit implied weights. Our results provide a unified perspective on sensitivity analysis across regression and weighting estimators and offer a tractable approach to sensitivity analysis in settings with heterogeneous treatment effects or more flexible estimators.Chapter 3 studies robust design and inference through covariate-adaptive randomization in matched observational studies. This is a published work with Sam Pimentel. This chapter develops randomization-based inference procedures that account for covariate imbalance and propensity score estimation uncertainty in matched designs. The chapter proposes covariate-adaptive randomization inference methods that adjust for residual covariate discrepancies after matching, while also incorporating sensitivity analysis for unobserved confounding. The resulting procedures improve finite-sample validity and robustness relative to standard randomization inference approaches. Through theoretical results, simulations, and empirical case studies, the chapter demonstrates how principled design-based adjustments can strengthen causal conclusions in observational studies.Taken together, the three chapters contribute to trustworthy causal inference in modern data settings. Across model screening, sensitivity analysis, and inference design, this dissertation emphasizes reality check, stability, and transparent uncertainty quantification as fundamental principles for trustworthy causal conclusions in the presence of unobserved variables, model uncertainty and residual covariate imbalance.