Bridging Multiple Worlds: Multi-marginal Optimal Transport for Causal Partial-identification Problem

Zijun Gao, Shu Ge, Jian QianAISTATS 2025 · OpenReview

Abstract

Under the prevalent potential outcome model in causal inference, each unit is associated with multiple potential outcomes but at most one of which is observed, leading to many causal quantities being only partially identified. The inherent missing data issue echoes the multi-marginal optimal transport (MOT) problem, where marginal distributions are known, but how the marginals couple to form the joint distribution is unavailable. In this paper, we cast the causal partial identification problem in the framework of MOT with $K$ margins and $d$-dimensional outcomes and obtain the exact partial identified set. In order to estimate the partial identified set via MOT, statistically, we establish a convergence rate of the plug-in MOT estimator for the $\ell_2$ cost function stemming from the variance minimization problem and prove it is minimax optimal for arbitrary $K$ and $d \le 4$. We also extend the convergence result to general quadratic objective functions. Numerically, we demonstrate the efficacy of our method over synthetic datasets and several real-world datasets where our proposal consistently outperforms the baseline by a significant margin (over 70%). In addition, we provide efficient off-the-shelf implementations of MOT with general objective functions.

Introduction

In causal inference, researchers face a fundamental challenge: for any given individual, we can only observe one potential outcome at a time. If a patient receives a treatment, we observe that specific result, but the counterfactual—what would have happened without that treatment—remains unobserved. Under the "potential outcome model," every individual has a set of potential outcomes for every possible treatment level. While we can estimate the average outcome for a group (marginal distributions), we lack the joint distribution that describes how these outcomes correlate across different treatment "worlds" for the same individual.

Because this joint distribution is missing, many critical causal quantities—such as the variance of a treatment effect—are partially identified. This means the data provides a range of plausible values (the identified set) rather than a single point estimate. Historically, researchers used conservative or naive bounds that were often too wide to be useful. This paper introduces a framework to find the "sharpest" (narrowest) possible bounds using the mathematics of optimal transport.

Methodology: MOT for Causal Inference

The authors bridge causal inference with Multi-marginal Optimal Transport (MOT). In this framework, the causal estimand is treated as a "cost function" and the observed marginal distributions of potential outcomes serve as "constraints."

Conceptual framework of Multi-marginal Optimal Transport (MOT) for causal partial identification.

The researchers utilize a "plug-in" estimator, solving the MOT problem using empirical distributions found in the data. Theoretically, the paper establishes a convergence rate of $O(n^{-1/2})$ for the MOT estimator when the outcome dimension $d \leq 4$. Remarkably, they show this rate is independent of the number of treatment levels ($K$), proving that their method is minimax optimal. To ensure practical utility, they implement efficient MOT solvers, including Sinkhorn and Greenkhorn algorithms.

Empirical Evaluation and Results

The approach was validated using synthetic data and three diverse real-world datasets: the Epitaxial Layer Growth engineering dataset, a Helpfulness social science dataset, and the STAR Project education dataset.

Performance improvement of MOT lower bounds over baseline across three real-world datasets.

As shown in the results, the MOT approach consistently outperformed baseline bounds by a significant margin. Specifically, the method achieved improvements of 102.8% on the Epitaxial dataset, 79.5% on the Helpfulness dataset, and 72.3% on the Education dataset. Furthermore, in the context of Neymanian confidence intervals, the proposal produced a variance estimate that was 8.16% lower than conventional conservative estimators, leading to much narrower and more useful confidence intervals.

Discussion and Future Work

By repurposing optimal transport as a tool for bounding missing information, this method provides the mathematically tightest possible intervals compatible with observed data. One of the most significant findings is that for low-dimensional outcomes ($d \leq 4$), the error in estimating the bound does not degrade as more treatment levels ($K$) are added, suggesting high scalability for complex experiments.

However, limitations remain. The convergence rate slows as the outcome dimension ($d$) increases, a manifestation of the "curse of dimensionality." Additionally, solving MOT problems is computationally intensive. Future research may explore low-rank tensor approximations to improve speed or incorporate covariates into a conditional MOT framework to yield even tighter, personalized causal bounds.

Summary updated

Abstract

Under the prevalent potential outcome model in causal inference, each unit is associated with multiple potential outcomes but at most one of which is observed, leading to many causal quantities being only partially identified. The inherent missing data issue echoes the multi-marginal optimal transport (MOT) problem, where marginal distributions are known, but how the marginals couple to form the joint distribution is unavailable. In this paper, we cast the causal partial identification problem in the framework of MOT with $K$ margins and $d$-dimensional outcomes and obtain the exact partial identified set. In order to estimate the partial identified set via MOT, statistically, we establish a convergence rate of the plug-in MOT estimator for the $\ell_2$ cost function stemming from the variance minimization problem and prove it is minimax optimal for arbitrary $K$ and $d \le 4$. We also extend the convergence result to general quadratic objective functions. Numerically, we demonstrate the efficacy of our method over synthetic datasets and several real-world datasets where our proposal consistently outperforms the baseline by a significant margin (over 70\%). In addition, we provide efficient off-the-shelf implementations of MOT with general objective functions.
BibTeX
@inproceedings{gao2025bridging,
  author = {Zijun Gao and Shu Ge and Jian Qian},
  title = {{Bridging Multiple Worlds: Multi-marginal Optimal Transport for Causal Partial-identification Problem}},
  booktitle = {International Conference on Artificial Intelligence and Statistics},
  year = {2025},
  url = {https://openreview.net/forum?id=h2WIiouJEm},
}
Figure