📊 statistics

Identifiability of Treatment Effects with Unobserved Spatially Varying Confounders

This paper establishes a general framework for the identifiability of treatment effects in linear models with unobserved spatially varying confounders, demonstrating that causal inference is feasible under mild conditions for various spatial models while also delineating scenarios where identifiability fails.

Original authors: Tommy Tang, Xinran Li, Bo Li

Published 2026-02-27

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Tommy Tang, Xinran Li, Bo Li

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

Imagine you are a detective trying to solve a crime: Did a specific treatment (like a new medicine or a policy) actually cause a change in health outcomes, or was it just a coincidence?

In the world of data science, this is called Causal Inference. Usually, the detective's biggest problem is the "Unseen Suspect"—a hidden factor that influences both the treatment and the outcome. For example, maybe people who live in a sunny neighborhood (the hidden factor) are more likely to exercise (the treatment) and also happen to be healthier (the outcome). If you don't account for the sun, you might wrongly blame the exercise for the health.

In standard detective work, if you can't see the suspect, you can't solve the case. But this paper argues that in spatial data (data tied to locations on a map), the "shape" of the neighborhood itself can act as a clue to catch the unseen suspect.

Here is the breakdown of the paper's findings using simple analogies:

1. The Core Problem: The "Ghost" in the Machine

The authors are studying a scenario where we have:

The Treatment ( $Z$ ): What we are testing (e.g., a new fertilizer).
The Outcome ( $Y$ ): The result (e.g., crop yield).
The Unseen Confounder ( $U$ ): A "Ghost" variable we can't measure (e.g., soil quality) that affects both the fertilizer choice and the crop yield.

If the Ghost is just random noise, we are stuck. But if the Ghost moves in a pattern across the map (e.g., soil quality changes gradually from north to south), the paper asks: Can we use the map's geometry to separate the Ghost from the Treatment?

2. The "Neighborhood" Analogy (Conditional Autoregressive Models)

The paper first looks at data where locations are neighbors (like houses on a street or counties on a map). They use a model called CAR (Conditional Autoregressive).

The Analogy: Imagine a row of houses. If your neighbor's house is painted blue, there's a good chance your house is blue too (spatial correlation).
The Discovery: The authors found that if the "Ghost" (unseen factor) has a specific spatial pattern, and the "Treatment" (exposure) has a different spatial pattern, we can mathematically untangle them.
The "Ring" vs. The "Web": Previous research said you needed a very specific, rigid structure (like a perfect ring of houses) to solve the mystery. This paper says, "No way!" You can solve it with almost any neighborhood layout, as long as it's not perfectly uniform.
- The Catch: If every house is connected to every other house in exactly the same way (a "fully connected" blob), the Ghost and the Treatment look identical, and the case remains unsolved. But if there are "indirect neighbors" (houses connected through a middleman but not directly), the map has enough "texture" to reveal the truth.

3. The "Spectral" Analogy (Leroux Models)

Next, they look at a more complex model (Leroux) that mixes different types of spatial patterns.

The Analogy: Think of the map as a musical chord. The "Ghost" plays a low note, and the "Treatment" plays a high note.
The Discovery: To tell the notes apart, the map needs to have enough "frequencies" (distinct distances or connections). If the map is too simple (like a single tone), you can't distinguish the notes. But if the map has enough variety (like a complex chord with at least 3 or 4 distinct notes), the math can isolate the "Treatment" note from the "Ghost" note.
The Warning: If the Ghost and the Treatment play the exact same note (same spatial pattern), you can never tell them apart. The paper proves exactly when this happens so researchers don't waste time trying to solve an impossible puzzle.

4. The "Linear Mix" Trap (Coregionalization)

The paper also tests a popular method called the Linear Model of Coregionalization.

The Analogy: Imagine the Ghost and the Treatment are two different colored lights shining on a wall, but they are both made by mixing the same three "base colors" (latent processes).
The Bad News: The authors prove that with this specific method, the case is unsolvable. No matter how much data you have, you can never separate the "Ghost" from the "Treatment." It's like trying to figure out how much red vs. blue is in a purple paint mixture when you only have the purple paint to look at.
The Lesson: If you use this specific model, your results might be "unidentifiable"—meaning your statistical software might give you an answer, but that answer could be completely wrong, and the computer won't tell you it's guessing.

5. The "Smoothness" Analogy (Matérn Models)

Finally, they look at continuous data (like temperature readings across a whole state) using Matérn models, which describe how "smooth" or "jagged" the data is.

The Analogy: Imagine the Ghost is a smooth rolling hill, and the Treatment is a bumpy, jagged mountain range.
The Discovery: If the "smoothness" of the Ghost is different from the "smoothness" of the Treatment, and you have data points far enough apart (like looking at the landscape from a plane), you can tell them apart.
The Condition: You need the "Ghost" and the "Treatment" to have different textures. If they are both equally smooth or equally jagged, you can't separate them.

The Big Takeaway

This paper is a rulebook for detectives.

Good News: You don't need perfect data or perfect maps. As long as your map has some "texture" (indirect neighbors, varied distances) and the hidden factor behaves differently than the treatment, you can find the truth.
Bad News: If the hidden factor and the treatment move in lockstep (perfectly correlated spatial patterns), or if you use the wrong mathematical model (like the Linear Coregionalization), you cannot find the truth.
The Warning: If you try to do causal inference in these "impossible" scenarios, your results will be unstable. It's like trying to weigh a feather on a scale that is broken; the number you get depends entirely on what you guess the weight should be, not the reality.

In short: The shape of your map holds the key to unlocking hidden causes, but only if the map isn't too simple and you don't use the wrong key.

1. Problem Statement

The paper addresses the challenge of causal inference in spatial settings where unmeasured confounders vary across space. In observational studies, unmeasured confounding ( $U$ ) that is correlated with the exposure ( $Z$ ) leads to biased estimates of the treatment effect ( $\beta$ ).

The Core Issue: Unlike independent data, spatial data possesses a correlation structure. If the spatial dependence of the unmeasured confounder is not properly modeled or if the model is not identifiable, the treatment effect cannot be uniquely determined from the observed data distribution.
The Gap: While recent literature (e.g., Schnell & Papadogeorgou, Guan et al.) has shown that treatment effects can be identified under specific spatial assumptions, a general framework for identifiability (the ability to uniquely recover parameters from data) across various spatial models and finite sample sizes was lacking.
Objective: To establish necessary and sufficient conditions under which the treatment effect $\beta$ is identifiable in linear spatial models with unobserved confounders, and to identify scenarios where identifiability fails.

2. Methodology

The authors propose a general linear structural model for $n$ spatial locations $s_{1:n}$ :
$Y = Z\beta + U + \epsilon$
where $Y$ is the outcome, $Z$ is the exposure, $U$ is the unmeasured spatial confounder, and $\epsilon$ is noise. The joint distribution of $(U, Z)$ is assumed to be Gaussian with a specific covariance structure $\Sigma$ .

The paper analyzes identifiability by examining whether the observed data distribution (the joint distribution of $Y$ and $Z$ ) uniquely determines the parameter $\beta$ . The authors decompose the problem into two main categories of spatial models:

A. Discrete/Areal Data (Conditional Autoregressive - CAR Models)

The authors study models where spatial dependence is defined via a proximity matrix $W$ (e.g., neighbor graphs).

Schnell & Papadogeorgou (S&P) CAR Model: A standard CAR structure where the precision matrix depends on $W$ .
Leroux CAR Model: A more flexible model allowing for different spatial dependence parameters for $U$ and $Z$ , including a parsimonious version where cross-covariance is constrained.

B. Continuous/Geostatistical Data

The authors study models defined over a continuous domain using distance matrices.

Linear Model of Coregionalization (LMC): $U$ and $Z$ are modeled as linear combinations of independent latent Gaussian processes.
Bivariate Spatial Covariance Models: General models where $U$ and $Z$ have specific covariance functions (e.g., Exponential, Gaussian, Matérn) with potentially different parameters.
Parsimonious Matérn Models: A specific constrained version of the Matérn covariance function often used in practice.

Technical Approach:

The authors utilize spectral decompositions of the spatial operators (e.g., $D-W$ for CAR models) to transform the problem into the frequency domain.
They derive conditions based on the eigenvalues of the spatial matrices and the linear independence of the covariance functions with respect to the observed distances.
They employ algebraic proofs to show that if two different sets of parameters yield the same observed data distribution, the treatment effect $\beta$ must be identical (identifiability) or can differ (non-identifiability).

3. Key Contributions and Results

A. Generalization of Identifiability in CAR Models

Finite Sample & General Graphs: The paper generalizes previous results (which required infinite samples and ring graphs) to finite sample sizes and general neighborhood structures.
Theorem 1 (S&P Model): Identifiability holds if:
1. The unmeasured confounder has spatial dependence ( $\phi_U \neq 0$ ).
2. The neighborhood graph is not "fully connected" in a specific sense (i.e., the proximity matrix $W$ has at least three distinct eigenvalues, or the degree matrix $D$ has distinct diagonal elements).
- Implication: If the graph is a fully connected clique (exchangeable correlation), $\beta$ is not identifiable.
Leroux Models (Theorems 2 & 3): Identifiability is established for both non-parsimonious and parsimonious models. Key conditions include:
- The spatial dependence structures of $U$ and $Z$ must be distinct (e.g., $\lambda_U \neq \lambda_Z$ ).
- If $\rho \neq 0$ (correlation exists), the cross-covariance structure must differ from the marginal structure of $Z$ .
- The matrix $D-W$ must have a sufficient number of distinct eigenvalues (at least 3 or 4 depending on the case).

B. Non-Identifiability of the Linear Model of Coregionalization (Theorem 4)

Critical Finding: The authors prove that under the standard Linear Model of Coregionalization (where $U$ and $Z$ share the same latent processes), the treatment effect $\beta$ is not identifiable, even if the true parameters of the latent processes are known.
Reason: The observed data only reveals the linear combination $\beta a_t + b_t$ for each latent process, making it impossible to separate the causal effect $\beta$ from the confounding coefficient $b_t$ .
Warning: This implies that statistical inference (especially Bayesian) using LMC is highly sensitive to priors and potentially unstable.

C. Identifiability in Geostatistical Models (Theorems 5 & 6)

Bivariate Models: Identifiability is achieved if the covariance functions for $U$ $U$ , $Z$ $Z$ , and their cross-covariance are K-linearly independent with respect to the observed distances.
- For common families (Exponential, Gaussian, Powered Exponential), this holds if the distance matrix contains at least 3 distinct positive values.
- The condition requires $\rho \neq 0$ and distinct spatial range parameters ( $\psi_{UZ} \neq \psi_Z$ ).
Parsimonious Matérn Models:
- Theorem 6: In an asymptotic setting (as $n \to \infty$ with distances diverging), identifiability holds if the smoothness parameters differ ( $\nu_U \neq \nu_Z$ ) and $\rho \neq 0$ .
- The authors conjecture finite-sample identifiability but note the complexity of Bessel functions makes a general proof difficult without large distances.

4. Significance and Implications

Theoretical Foundation: The paper provides the first rigorous, general framework for determining when causal effects are identifiable in the presence of spatial confounding. It moves beyond "assuming" identifiability to providing verifiable conditions based on the data geometry (neighborhood structure or distance matrix).
Practical Guidelines for Researchers:
- Check the Graph: For areal data, researchers must verify that their spatial graph is not fully connected (e.g., a ring or grid with holes is better than a fully connected clique).
- Model Selection: The paper explicitly warns against using the Linear Model of Coregionalization for causal inference, as it leads to non-identifiability. Instead, models with distinct spatial dependence structures for exposure and confounder (like the Leroux model or specific bivariate covariance models) are recommended.
- Data Requirements: Identifiability often relies on the diversity of spatial scales (distinct eigenvalues or distances). Sparse or highly regular data may fail to support identification.
Diagnostic Tools: The authors provide observable quantities (e.g., checking if $E(YZ^T)E(ZZ^T)^{-1}$ is proportional to the identity matrix) that researchers can compute from data to assess whether identification conditions (like $\rho \neq 0$ ) are likely met.
Cautionary Note: The paper highlights that in non-identifiable scenarios, Bayesian inference is not a "fix" but rather shifts the burden to the prior, which can lead to misleading conclusions if the prior is not carefully justified.

Conclusion

This paper fundamentally advances spatial causal inference by delineating the boundary between identifiable and non-identifiable scenarios. It demonstrates that while spatial structure offers a mechanism to control for unmeasured confounding, this mechanism only works if the spatial dependence of the confounder and exposure is sufficiently distinct and the data geometry supports the separation of these effects. The results serve as a critical guide for model selection and study design in environmental science, epidemiology, and other fields utilizing spatial data.