A1827
Title: Causal partial identification via optimal transport
Authors: Zijun Gao - University of Southern California (United States) [presenting]
Abstract: In causal inference, only one of multiple potential outcomes is observed for each unit, leaving many causal quantities only partially identified (PI). This missingness closely parallels the optimal transport (OT) problem, where marginal distributions are observed but the joint coupling between them is unknown. This connection is formalized by casting the causal partial identification problem within the OT framework, enabling PI sets to be analyzed using tools from the rapidly developing optimal transport literature. When treatment effects are heterogeneous, incorporating covariate information can further sharpen PI sets. The OT formulation is generalized to a conditional optimal transport (COT) framework and new statistical tools for COT are developed to extend the analysis of PI sets. Real-world applications and future research directions are discussed.
