Empirical Exploration of the Explanation-Guided Learning for Causal Effect Estimation

This thesis investigates the integration of Explanation-Guided Learning (EGL) into causal inference, with a focus on Conditional Average Treatment Effect (CATE) estimation from observational data. While EGL has been widely studied in areas such as computer vision and natural language processing with an objective to improve explainability and generalizability of models, its application in the domain of causal inference remains largely unexplored. We propose a framework that incorporates EGL, specifically with counterfactual explanations, into modern CATE estimators: Meta-Learners. We improve accuracy of various meta-learning architectures by integrating explanation-based loss following the latest practices of the EGL literature. The proposed method enables the integration of structured domain knowledge into causal learning pipelines, which to the best of our knowledge, has not been done before. To evaluate the effectiveness of this approach, we conduct controlled experiments using synthetic data with known ground-truth treatment effects and varying levels of confounding. We assess performance across several meta-learners, including S-, X-, DR-, and domain adaptation variants. Our results demonstrate that incorporating explanation signals improves the accuracy of CATE estimation. In particular, the DA-learner and S-learner show the greatest reductions in mean absolute bias, with decreases of 49% and 51.8%, respectively. Overall, this work bridges two previously disconnected research areas – a newly emerged EGL and causal inference on observational data – and provides experimental proofs that demonstrate how explanation signals improve the accuracy of CATE estimation.