📊 statistics

Spatial Causal Tensor Completion for Multiple Exposures and Outcomes: An Application to the Health Effects of PFAS Pollution

This paper proposes a spatial causal tensor completion framework that jointly models multiple exposures and outcomes while adjusting for unmeasured spatial confounders and missing data, demonstrating its effectiveness in estimating the causal health effects of PFAS mixtures through theoretical guarantees, simulations, and a national application.

Original authors: Xiaodan Zhou, Brian J Reich, Shu Yang

Published 2026-03-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Xiaodan Zhou, Brian J Reich, Shu Yang

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 Big Picture: The "Forever Chemical" Mystery

Imagine the environment is a giant, complex soup. In this soup, we have PFAS (Per- and polyfluoroalkyl substances), often called "forever chemicals" because they never seem to break down. These chemicals are everywhere in our drinking water, but they rarely show up alone. Usually, you have a mix of different types (like PFOA and PFOS) swirling together.

Scientists want to know: Does drinking water with these chemicals cause specific diseases like asthma, obesity, or high blood pressure?

The problem is that the world is messy. People who drink this water also live in specific neighborhoods. Those neighborhoods might have other problems (like older housing, different jobs, or traffic pollution) that also cause health issues. If you just look at the data, it's hard to tell if the sickness is caused by the chemical or by the neighborhood. This is called confounding.

Furthermore, we don't have data for every single person. We have data for some water systems but not others, and we have data for some diseases but not all. It's like trying to finish a giant jigsaw puzzle where half the pieces are missing and the picture on the box is blurry.

The Solution: A "Smart 3D Puzzle Solver"

The authors (Zhou, Reich, and Yang) built a new statistical tool called Spatial Causal Tensor Completion. That's a mouthful, so let's break it down with a metaphor.

1. The "Tensor" (The 3D Spreadsheet)

Imagine a standard spreadsheet (a 2D table) with rows and columns.

Rows: Different towns (Public Water Systems).
Columns: Different chemicals (PFOA, PFOS).
Layers: Different diseases (Asthma, Heart Disease, etc.).

Now, imagine stacking those spreadsheets on top of each other to make a 3D cube. This is a "Tensor." It holds all the information about who got sick, where they lived, and what chemicals were in their water.

2. The "Missing Pieces" (Tensor Completion)

In this 3D cube, many squares are empty (gray). We know a town has PFOA, but we don't know if they have PFOS. We know they have asthma data, but not kidney disease data.

Old way: Try to guess the missing pieces one by one, ignoring the rest of the puzzle.
New way: The authors assume the puzzle has a hidden, simple pattern (a "low-rank" structure). Just like a painting might be made of only a few main colors, the health data is driven by a few main underlying factors. Their algorithm looks at the pieces we do have and uses those patterns to intelligently fill in the missing pieces.

3. The "Ghost in the Machine" (Spatial Confounding)

Here is the tricky part. Some missing pieces aren't just random; they are missing because of geography.

Analogy: Imagine you are trying to figure out if eating spicy food causes heartburn. But you notice that everyone who eats spicy food also lives in a hot, humid climate. Maybe the humidity is actually what's causing the heartburn, not the food!
In this study, the "humidity" is unmeasured spatial confounding. Maybe a town has high chemical levels and high disease rates because of an old factory nearby that we didn't measure.
The Fix: The authors use a "spectral adjustment." Think of this as a noise-canceling headphone for maps. They use math (Graph Laplacian eigenvectors) to identify the "hum" of the geography—the smooth, rolling hills of disease and pollution that aren't caused by the chemicals. They subtract this "geographic hum" from the data so they can hear the true signal of the chemicals.

How They Did It (The Recipe)

They used a three-step cooking process:

Listen to the Map: First, they looked at the map to find the "geographic hum" (the unmeasured factors) and made a rough guess at the missing data.
Check the Odds: They calculated how likely a town was to have these chemicals based on its demographics (income, jobs, etc.). This helps them weigh the data so that towns with rare chemical mixes don't skew the results.
Refine the Puzzle: They put the weights and the geographic corrections back into the 3D puzzle solver. This time, the solver fills in the missing pieces while ignoring the geographic noise.

What They Found (The Results)

When they applied this new method to data from 5,495 water systems across the US:

The Old Way (Standard Methods): The old methods screamed, "Everything is bad!" They found that almost every chemical mix caused almost every disease. The odds ratios were huge (like 3.0 or 4.0), suggesting massive risks.
The New Way (Spatial Tensor): The new method said, "Whoa, let's calm down." After removing the geographic noise and filling in the missing data smartly, most of those scary links disappeared.
- PFOA: Showed almost no link to diseases (except maybe a tiny protective effect against asthma, which is likely just a fluke).
- PFOS: Still showed a real, small, but significant link to four things: High blood pressure, tooth loss, asthma, and obesity.

The Takeaway

The paper teaches us a vital lesson: Don't trust the first map you see.

When studying environmental health, geography is a master of disguise. If you don't account for the fact that sick people often live in specific types of neighborhoods, you might blame the wrong thing.

By using this "3D puzzle solver" that understands how geography works, the authors found that while PFOS is indeed a problem for specific health issues (like blood pressure and teeth), the panic over every chemical causing every disease was likely an exaggeration caused by bad math.

In short: They built a smarter way to look at the data, filtered out the "geographic static," and found that the truth is more nuanced and less chaotic than we thought.

1. Problem Statement

The paper addresses the challenge of estimating joint causal effects of multiple chemical exposures (mixtures) on multiple health outcomes using spatially dependent observational data. Specifically, the authors focus on Per- and polyfluoroalkyl substances (PFAS), which are ubiquitous environmental contaminants often found as mixtures (e.g., PFOA and PFOS) in drinking water.

Key Challenges:

Mixture Complexity: PFAS chemicals rarely occur in isolation; isolating the effect of individual components from a mixture is difficult.
Multiple Endpoints: PFAS exposure is linked to a wide range of outcomes (e.g., kidney disease, cancer, hypertension), requiring a multivariate approach rather than univariate analysis.
Unmeasured Spatial Confounding: Both exposure levels and health outcomes exhibit spatial patterns. Unmeasured variables (e.g., industrial activity, socioeconomic trends) that vary smoothly over space can confound the relationship between exposure and outcome.
Missing Potential Outcomes: In observational studies, only one outcome is observed per unit (the realized exposure), leaving a vast number of counterfactual outcomes missing.

2. Methodology: Spatial Causal Tensor Completion

The authors propose a unified framework that combines low-rank tensor completion with spectral adjustment for spatial confounding.

A. Data Structure and Formulation

Tensor Representation: The data is organized into a 3-way tensor $\mathcal{Y} \in \mathbb{R}^{N \times L \times O}$ $Y \in R^{N \times L \times O}$ , where:
- $N$ : Number of spatial units (Public Water Systems).
- $L$ : Number of exposure combinations (e.g., None, PFOA only, PFOS only, Both).
- $O$ : Number of health outcomes.
Low-Rank Assumption: The potential outcome tensor $\mathcal{Y}^*$ $Y^{*}$ is assumed to follow a Tucker decomposition:
$\mathcal{Y}^* = \mathcal{G} \times_1 \mathbf{U}_1 \times_2 \mathbf{U}_2 \times_3 \mathbf{U}_3 + \mathcal{E}$
- $\mathcal{G}$ : Core tensor capturing interactions.
- $\mathbf{U}_1, \mathbf{U}_2, \mathbf{U}_3$ : Factor matrices representing spatial, exposure, and outcome variations, respectively.
- Spatial Factor ( $\mathbf{U}_1$ ): Modeled as a combination of measured covariates ( $\mathbf{Z}$ ) and unmeasured latent spatial components ( $\mathbf{S}$ ).

B. Handling Unmeasured Spatial Confounding

To address unmeasured spatial confounders ( $\mathbf{S}$ ), the method employs spectral adjustment:

A graph Laplacian is constructed based on spatial adjacency.
The unmeasured spatial component is approximated by the span of the first $k$ low-frequency eigenvectors of the Laplacian ( $\mathbf{S} \approx \Phi_{1:k} \beta$ ).
This assumes that unmeasured confounders vary smoothly over space and can be captured by these dominant spatial trends.

C. Estimation Algorithm (Three-Step Procedure)

The estimation is implemented via a Spatial Projected Gradient Descent (S-PGD) algorithm:

Step 1 (Spectral Adjustment): Fit an unweighted Tucker model using the Laplacian eigenvectors to obtain an initial estimate of the latent spatial component ( $\hat{\mathbf{S}}^{(0)}$ ).
Step 2 (Propensity Score): Estimate exposure probabilities $\pi(\ell | \mathbf{Z}_i, \hat{\mathbf{S}}^{(0)}_i)$ using multinomial logistic regression and construct inverse probability weights (IPW) to create a pseudo-population where exposure is independent of potential outcomes.
Step 3 (Weighted Completion): Re-fit the Tucker model using the IPW weights to correct for exposure-induced selection bias, yielding the final completed potential outcome tensor $\hat{\mathcal{Y}}$ and refined spatial component $\hat{\mathbf{S}}$ .

D. Causal Inference

Average Treatment Effect (ATE): Calculated by comparing imputed potential outcomes.
Estimator: The authors use an Augmented Inverse Probability Weighting (AIPW) estimator. This estimator is doubly robust, achieving $\sqrt{N}$ -consistency if either the outcome model (tensor) or the propensity score model is correctly specified.
Theoretical Guarantees: The paper establishes Frobenius error bounds for the estimated tensor, showing that the error depends on spatial approximation error, propensity score estimation error, and tensor rank.

3. Key Contributions

Unified Framework: Development of a causal inference framework that jointly models multiple binary exposures and multiple outcomes using low-rank tensor completion, moving beyond single-outcome, single-exposure analyses.
Spatial Confounding Integration: Novel integration of graph-Laplacian spectral adjustment directly into the tensor completion procedure to approximate and adjust for unmeasured spatial confounders.
Theoretical Rigor: Establishment of rigorous theoretical guarantees, including Frobenius error bounds and conditions for $\sqrt{N}$ -consistency of the ATE estimator.
Algorithmic Innovation: Implementation of a spatial projected-gradient descent algorithm that adaptively selects eigenvectors to capture spatial structure without pre-specifying the exact number of components.

4. Results

The method was applied to national PFAS monitoring data covering 5,495 Public Water Systems across 47 states and 13 chronic disease outcomes.

Comparison with Alternatives: The proposed "Spatial-Tensor" method was compared against standard regression, non-spatial tensor completion, and spatial propensity score methods.
- Regression-based methods produced inflated, implausible odds ratios (often > 3.0), likely due to unadjusted spatial confounding.
- Non-spatial tensor methods identified many significant adverse associations but failed to account for geographic artifacts.
- Spatial-Tensor Method: Produced significantly more conservative and credible estimates. After adjusting for spatial confounding, most previously reported associations were attenuated.
Specific Findings:
- PFOA: Showed no significant adverse associations with any outcome after adjustment (except a modest protective effect on asthma).
- PFOS: Retained significant adverse associations with hypertension, tooth loss, asthma, and obesity. These associations persisted after spatial adjustment, suggesting they are less likely to be geographic artifacts.
Latent Structure: The model successfully identified distinct spatial clusters (e.g., Eastern US, South-Central) and shared disease pathways, demonstrating the utility of the low-rank structure in capturing complex data patterns.

5. Significance

Environmental Health: The study underscores that failing to account for spatial confounding can lead to substantial overestimation of health risks in environmental epidemiology. It provides a more rigorous basis for regulatory decisions regarding PFAS.
Methodological Advancement: It bridges the gap between tensor completion (often used in recommendation systems) and causal inference, offering a scalable solution for high-dimensional mixture problems with spatial dependencies.
Policy Implications: By yielding more conservative estimates, the method suggests that while PFOS remains a concern for specific outcomes (hypertension, obesity), the broad spectrum of associations reported in previous literature may be inflated by unmeasured geographic factors.

In summary, this paper presents a robust statistical solution for disentangling the complex, spatially correlated effects of chemical mixtures on public health, offering a template for future causal inference in environmental science.