🤖 AI

A primer on optimal transport for causal inference with observational data

This review paper introduces the deep, often overlooked connections between optimal transport theory and causal inference with observational data, demonstrating how optimal transport principles underpin foundational causal models and aiming to unify terminology across disciplines while identifying new research directions.

Original authors: Florian F Gunsilius

Published 2026-08-10

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Florian F Gunsilius

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Great Detective Game: Uncovering Hidden Causes

Imagine you are a detective trying to solve a mystery, but with a twist: you can only look at the crime scene after the crime has happened, and you never got to see the suspect's actions in real time. This is the daily struggle of scientists trying to understand causality—figuring out if one thing (like taking a new medicine or building a factory) actually causes another thing (like getting better or changing the local economy). The problem is that we can never see two versions of reality at once. We can't see a person who took a medicine and the exact same person who didn't take it, side-by-side, to see who got sick. We only get to see one path.

To solve this, statisticians have developed a toolkit to compare different groups of people, hoping to find a "control group" that acts like a perfect twin for the people who received the treatment. But often, these groups aren't perfect twins; they have hidden differences (like motivation or genetics) that mess up the comparison. For decades, researchers have used clever math tricks to fix this, but they often didn't realize they were all using the same secret weapon. That weapon is called Optimal Transport. Think of it as the ultimate logistics puzzle: imagine you have a pile of sand in one spot and a pile of sand in another, and you want to move the first pile to perfectly match the shape of the second pile using the least amount of energy. This paper reveals that the math used to move sand is actually the hidden foundation for many of the most important methods we use to figure out cause-and-effect in the real world.

The Paper's Big Reveal: The Hidden Map

In this paper, Florian Gunsilius acts as a guide, showing us that the complex world of causal inference (figuring out what causes what) and optimal transport (the math of moving probability distributions efficiently) are actually old friends who have been speaking different languages for years. The author argues that many of the standard tools economists and statisticians have used for decades to analyze observational data—data where we can't run controlled experiments—are secretly built on the principles of optimal transport, even if the researchers didn't know it.

The paper doesn't just point out this connection; it suggests that realizing this link can help us build better, more powerful tools. Here is what the author finds and explains:

1. The "Monotone Rearrangement" is the Secret Sauce
The paper starts by explaining a concept called the monotone rearrangement. Imagine you have a deck of cards representing people with different levels of "hidden ability" (an unobservable trait), and you want to see how that ability translates into "earnings." If you assume that higher ability always leads to higher earnings (a rule called monotonicity), you can simply line up the lowest-ability person with the lowest earner, the second-lowest with the second-lowest, and so on. This creates a perfect map between the hidden trait and the outcome.
The author shows that this simple "lining up" is actually a specific type of optimal transport problem. This insight is powerful because it allows researchers to identify not just the average effect of a treatment, but the entire story of how different people are affected. For example, in a famous debate about minimum wage, some studies said raising wages hurt jobs, while others said it helped. By using this "optimal transport" view, researchers found that the truth was in the middle: raising wages helped small restaurants but hurt large ones. The average was hiding the real story, but the transport map revealed it.

2. Fixing the "Bad Twin" Problem with Instruments
Sometimes, you can't find a perfect control group because the people choosing the treatment are different from those who don't (this is called endogeneity). To fix this, researchers use instrumental variables—a third factor that influences who gets the treatment but doesn't directly affect the outcome.
The paper explains that the math used to make these instruments work is also based on optimal transport. Specifically, it introduces a method called fixed-point iteration. Imagine you are trying to find a meeting spot between two groups. You start at a point, move toward the other group, then move back, and keep repeating this "ping-pong" motion until you settle on a spot where the groups align. The paper shows that this mathematical "ping-pong" is actually a dynamic system of moving probability distributions (like a Brenier map). This helps researchers identify causal effects even when the data is messy, though the author notes that this gets very complicated when dealing with multiple variables at once.

3. Building Synthetic Twins with "Sand Piles"
Another popular method is Synthetic Controls, where researchers build a fake "control unit" by mixing together several real ones (like creating a fake state by mixing parts of California, Texas, and New York) to see what would have happened if a treatment hadn't occurred.
The paper connects this to Wasserstein barycenters, which is a fancy way of saying "the average of a pile of sand." Instead of just averaging numbers, this method averages entire distributions (shapes). The author suggests that by using this transport-based approach, we can create "distributional synthetic controls" that match the entire shape of the data, not just the average. This is useful when the people in the groups change over time (like restaurants opening and closing), which standard methods struggle with.

4. When Twins Don't Match: Unbalanced Transport
Finally, the paper tackles a common problem: what if the treatment group and the control group are so different that you can't find a match for everyone? Standard "optimal transport" requires every grain of sand in one pile to be moved to the other, which forces researchers to match people who aren't actually similar, leading to bad guesses.
The author proposes using unbalanced optimal transport. Think of this as a logistics system that allows you to leave some sand behind or add a little extra if the piles don't match perfectly. This method automatically discards the people who don't have a good match, rather than forcing a bad match. The paper suggests this could lead to more accurate results in real-world studies where perfect matches are rare, though it notes that more research is needed to prove exactly how much better it is compared to old methods.

What the Paper Says We Should Be Careful About

The author is careful to point out that while these connections are exciting, they aren't magic wands.

It's not a solved problem: The paper suggests these methods are promising and offers new ways to think about old problems, but it doesn't claim that optimal transport solves every causal inference issue. In fact, it highlights that some extensions (like moving from one dimension to many) are mathematically tricky and not always straightforward.
Assumptions still matter: The "monotone rearrangement" (the idea that higher ability always means higher earnings) is a strong assumption. The paper notes that if this assumption is wrong, the map breaks. It suggests that by viewing these problems as transport problems, we can test different "cost functions" (different rules for how we move the sand) to see if our conclusions are robust.
Data limitations: The paper mentions that some of the most advanced methods (like the "fixed-point" iterations for instrumental variables) rely on specific mathematical conditions, like having a continuous range of data. If the data is too sparse or "chunky," these fancy methods might not work as well.

The Takeaway

Ultimately, this paper is a unifying map. It tells us that the tools statisticians have been using to solve the "fundamental problem of causal inference" (we can't see two worlds at once) have been secretly using the physics of moving probability distributions all along. By making this connection explicit, the author hopes to unify the language between different fields—economics, statistics, and machine learning—and open the door to new, more flexible ways of understanding how the world works. It's a reminder that sometimes, the most powerful insights come from realizing that two different puzzles are actually pieces of the same picture.

Technical Summary: A Primer on Optimal Transport for Causal Inference with Observational Data

Problem Statement
The paper addresses the fundamental challenge in causal inference with observational data: the inability to observe counterfactual states simultaneously (the "fundamental problem of causal inference"). In observational settings, treatment assignment is endogenous, meaning the treatment variable $X$ is correlated with unobserved confounders $U$ that also affect the outcome $Y$ . This creates a "back-door path" ( $U \to X$ ) that biases standard estimation. While traditional methods (e.g., instrumental variables, difference-in-differences, matching) have been developed to address this, the author argues that the underlying mathematical structures of these methods have implicitly relied on principles of Optimal Transport (OT) for decades without explicit recognition. The paper seeks to make these connections explicit, arguing that OT is not merely a tool but the foundational framework for identifying causal mechanisms, particularly when accounting for treatment heterogeneity.

Methodology and Theoretical Framework
The review systematically connects structural causal models to OT theory, moving from univariate to multivariate settings and from point identification to partial identification.

Structural Models and Monotone Rearrangement:
The paper begins with the structural model $Y = g(X, U)$ , where $g$ is the causal mechanism and $U$ represents unobserved heterogeneity. In exogenous settings ( $X \perp U$ ), the identification of $g$ relies on the monotone rearrangement. If $g(x, \cdot)$ is strictly monotone, it acts as the optimal transport map pushing the distribution of unobservables $P_U$ to the conditional distribution of outcomes $P_{Y|X=x}$ . This map is the solution to the Monge problem with a convex cost function, specifically the quantile function $g(x, u) = F^{-1}_{Y|X=x}(F_U(u))$ . This establishes that identifying the causal mechanism is equivalent to finding the optimal coupling between unobservables and outcomes.
Instrumental Variables (IV) and Control Variables:
In endogenous settings ( $X \not\perp U$ ), the paper analyzes Instrumental Variable (IV) models ( $Y=g(X,V), X=h(Z,W)$ ).
- Control Variables: It demonstrates that the classic control variable approach (e.g., [87]) is based on the monotone rearrangement. The control variable $R = F_{X|Z}(X)$ mimics the unobservable $W$ . The paper generalizes this by showing that any measure-preserving isomorphism can serve as a valid control variable, extending the method beyond uniform distributions.
- Fixed-Point Iterations: For binary instruments, the paper reviews identification strategies using fixed-point iterations (e.g., [141, 49]). These methods utilize "vertical" (changing $Z$ ) and "horizontal" (changing quantiles) maps to trace the causal mechanism $g$ back to a fixed point where $X$ is effectively exogenous.
- Multivariate Extensions: The paper discusses extending these ideas to multivariate settings using Brenier maps (gradients of convex functions). It highlights that while Brenier maps provide a natural generalization of monotonicity via cyclic monotonicity, they impose strict dimensionality constraints (the unobservable $W$ must match the dimension of $X$ ), which limits their applicability in some causal contexts.
- Partial Identification: For settings where full identification is impossible, the paper frames the problem of bounding average treatment effects as an optimal transport problem on path spaces. This involves optimizing over joint distributions of counterfactual processes subject to observed marginal constraints, linking causal bounds to problems in statistical physics (e.g., large particle dynamics).
Difference-in-Differences (DiD) and Synthetic Controls:
- Changes-in-Changes: The paper reviews the nonparametric DiD estimator (e.g., [11]), showing it relies on the monotone rearrangement to transport the "natural trend" of the control group to the treated group.
- Multivariate DiD: It extends this to multivariate outcomes using cyclic comonotonicity, allowing for the identification of counterfactual distributions via Brenier maps without requiring individual production functions to be monotone, only that they are cyclically comonotone.
- Distributional Synthetic Controls: The paper introduces a novel extension of synthetic controls where the goal is to match the distribution of outcomes rather than just the mean. This is formulated as finding a Wasserstein barycenter of control units to replicate the treated unit's pre-treatment distribution. The identification relies on the assumption that the transport maps between time periods are isometries in the Wasserstein space.
Matching and Unbalanced OT:
The paper critiques standard matching methods for their reliance on the strict mass-preserving constraint of classical OT, which forces matches even when covariate supports do not overlap (violating the overlap assumption). It proposes Unbalanced Optimal Transport as a solution. By relaxing the marginal constraints and adding divergence penalties (e.g., Kullback-Leibler or $\phi$ -divergence), unbalanced OT allows for "partial matching," automatically discarding units without good matches. This provides a disciplined way to handle violations of the overlap assumption and reduces bias in finite samples.

Key Contributions and Results

Unification of Notation and Theory: The paper unifies the language of econometrics, statistics, and mathematics by revealing that foundational causal models (structural equations, DiD, IV, synthetic controls) are implicitly built on OT principles.
Generalization of Control Variables: It provides a rigorous generalization of the control variable approach, showing that the specific construction of the control variable can be any measure-preserving isomorphism, not just the cumulative distribution function.
Dynamic Identification: It elucidates the connection between causal identification in IV models and gradient dynamics (specifically, the inverse Brenier map acting as a gradient ascent flow), offering new mathematical perspectives on convergence and identification.
Path Space OT: It frames partial identification in IV models as an optimal transport problem on path spaces, bridging causal inference with statistical physics and large deviation theory.
Distributional Synthetic Controls: It introduces a distributional version of synthetic controls based on Wasserstein barycenters, allowing for the analysis of heterogeneous effects at the aggregate level.
Unbalanced Matching: It advocates for unbalanced OT in matching problems to naturally handle support violations, offering a theoretically grounded alternative to ad-hoc trimming or nearest-neighbor matching.

Significance and Claims
The author claims that recognizing the OT foundation of causal inference is crucial for several reasons:

Clarification of Assumptions: It clarifies that assumptions like monotonicity are specific instances of choosing a cost function in an OT problem. This opens the door to sensitivity analyses by varying the cost function (e.g., using concave costs for non-monotone mechanisms).
Generalization: It provides a pathway to generalize existing methods to higher dimensions and more complex settings (e.g., multivariate DiD, distributional synthetic controls) by leveraging the rich mathematical theory of OT.
Cross-Disciplinary Bridge: The paper positions OT as a bridge between causal inference and other fields, such as statistical physics (via path space formulations) and machine learning (via distributional matching).
Practical Utility: By making these connections explicit, the review aims to unify disparate literatures, allowing researchers to transfer tools and insights (e.g., entropic regularization for computational efficiency, unbalanced OT for robustness) across different causal inference problems.

The paper concludes that while the connection between OT and causal inference has existed implicitly, making it explicit allows for the development of more robust, general, and interpretable methods for identifying causal effects in observational data.

The Great Detective Game: Uncovering Hidden Causes

The Paper's Big Reveal: The Hidden Map

What the Paper Says We Should Be Careful About

The Takeaway

Technical Summary: A Primer on Optimal Transport for Causal Inference with Observational Data

More like this