Causal Optimal Transport for Treatment Effect Estimation.

Area of Science:

• Causal inference
• Machine learning
• Optimal transport theory

Background:

• Estimating treatment effects is crucial for understanding intervention impacts.
• Treatment assignment bias and limited covariate overlap challenge classical causal inference.
• Supervised methods struggle with counterfactual inference due to factual space overfitting.

Purpose of the Study:

• To propose a novel causal optimal transport (CausalOT) model for accurate individual treatment effect (ITE) estimation.
• To address limitations of propensity score methods and supervised approaches in causal inference.
• To leverage global covariate information for improved counterfactual outcome prediction.

Main Methods:

• Developed the Causal Optimal Transport (CausalOT) model based on optimal transport theory.
• Introduced a novel propensity measure and a regularized optimal transport problem.
• Designed a counterfactual loss function to align factual and counterfactual outcome distributions.
• Proved the theoretical generalization bound for CausalOT's counterfactual error.

Main Results:

• CausalOT effectively utilizes global covariate information to mitigate issues from limited overlapping units.
• The model demonstrates improved counterfactual inference by aligning factual and counterfactual outcomes.
• Empirical studies on benchmark datasets show CausalOT surpasses state-of-the-art causal inference methods.

Conclusions:

• CausalOT offers a robust framework for individual treatment effect estimation.
• The method enhances causal inference by overcoming limitations in existing approaches.
• CausalOT shows significant promise for real-world applications requiring precise treatment effect evaluation.