The ubiquity of decision making has sparked a surge of research in treatment effects estimation across disciplines such as statistics, healthcare, economics, and financial applications. With the growing prominence of data science, causal machine learning has become appealing for inferring treatment effects in observational studies, surpassing the limitations of classical machine learning methods and randomized controlled trials (RCTs). Classical machine learning methods, despite their notable predictive capabilities, are constrained to analyzing data correlations rather than establishing causal relationships such as treatment effects, especially when data exhibit strong selection bias. Besides, decision-makers do not have the luxury to conduct RCTs due to their costliness and time-consuming nature. This thesis will introduce two distinct methodologies that are designed to estimate Average Treatment Effect (ATE) and Individual Treatment Effect (ITE) separately.

The Double/Debiased Machine Learning (DML) is one of the prevalent methods to estimate ATE in observational studies. However, the DML estimators can suffer from an error-compounding issue and even give an extreme estimate when the propensity scores are misspecified or very close to 0 or 1. Previous studies have solved this problem through some empirical tricks such as propensity score trimming, yet none of the existing works solves this problem from a theoretical standpoint. In this paper, we propose a Robust Causal Learning (RCL) method to offset the deficiencies of the DML estimators. Theoretically, the RCL estimators i) are as consistent and doubly robust as the DML estimators, and ii) can alleviate the error-compounding issue. Empirically, the comprehensive experiments show that i) the RCL estimators give more stable estimations of the causal parameters than the DML estimators, and ii) the RCL estimators outperform the traditional estimators and their variants when applying different machine learning models on both simulation and benchmark datasets.

The representation balancing methods are recent advanced neural network models that attempt to estimate ITE by leveraging covariate balancing in the representation space. The rationale behind these methods is that counterfactual estimation relies on (1) preserving the predictive power of factual outcomes and (2) balanced representations between treated and controlled groups. However, a trade-off naturally occurs when achieving these two objectives simultaneously, as having domain-invariant representations can be accompanied by a failure of domain discrimination, resulting in the loss of domain-related information, especially when representations encounter an over-balancing issue. In this thesis, we propose a novel framework, doubly robust representation balancing (DRRB), to deal with the over-balancing issue. Simultaneously, we also design an outcome balancing (OB) regularization to force the model subject to the unconfoundedness assumption to improve ITE estimation. The complete model is referred to as DRRB-OB. The extensive experimental results on benchmark datasets and a newly introduced credit dataset show a general outperformance of our method compared with existing methods.