Robust Spatial Confounding Adjustment via Basis Voting
Anik Burman, Elizabeth L. Ogburn, Abhirup Datta Department of Biostatistics, Johns Hopkins University · 2025
Read Robust Spatial Confounding Adjustment via Basis Voting by Anik Burman, Elizabeth L. Ogburn, Abhirup Datta Department of Biostatistics, Johns Hopkins University (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-22464.txt.
Introduction
Anik Burman and two co-authors address spatial confounding, where an unmeasured factor that varies across space can distort the estimated effect of an exposure. Standard methods assume the exposure changes at finer scales than the confounder, an assumption that may fail. Their framework relaxes that demand and instead needs only a few basis shapes tied to the exposure but not the confounder.
Their estimator, basis voting, builds many candidate estimates, each from one basis shape, and takes the most common answer. They also supply a sharper candidate estimate than the usual projection based one. Simulations and theory support the approach for spatial health and policy studies.
Catalog record
| CREATOR | Anik Burman, Elizabeth L. Ogburn, Abhirup Datta Department of Biostatistics, Johns Hopkins University |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 105 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2510-22464
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Robust Spatial Confounding Adjustment via Basis Voting Anik Burman, Elizabeth L. Ogburn, Abhirup Datta Department of Biostatistics, Johns Hopkins University Abstract Estimating causal effects of spatially structured exposures is complicated by unmeasured spatial confounders, which undermine identifiability in spatial linear regression models unless structural assumptions are imposed. We develop a general framework for causal effect estimation that relaxes the commonly assumed requirement that exposures contain higher-frequency variation than confounders. We propose basis voting , a plurality-rule estimator — novel in the spatial literature — that consistently identifies causal effects only under the assumption that, in a spatial basis expansion of the exposure and confounder, there exist several basis functions in the support of the exposure but not the confounder. This assumption generalizes existing assumptions of differential basis support used for identification of the causal effect under spatial confounding, and does not require prior knowledge of which basis functions satisfy this support condition. We also show that the standard projection-based estimator used in other methods relying on differential support is inefficient, and provide a more efficient novel estimator. Extensive simulations and a real-world application demonstrate that our approach reliably recovers unbiased causal estimates whenever exposure and confounder signals are separable on a plurality of basis functions. Importantly, by not relying on higher-frequency variation, our method remains applicable to settings where exposures are smooth spatial functions, such as distance to pollution sources or major roadways, common in environmental studies. Keywords: spatial statistics; geostatistics; plurality rule; mode estimation; kernel density; environmental exposures 1 Introduction Estimating the true association between spatially structured exposures and outcomes is a fundamental problem in spatial statistics and environmental epidemiology. However, this task is challenging in the presence of unmeasured spatial variables that simultaneously influence both the exposure and the outcome. When such variables are not accounted for, the (conditional) effect size in standard regression models applied to point-referenced spatial data is unidentifiable, leading to biased estimates. This phenomenon is referred to as spatial confounding , and it arises frequently in real-world applications where spatial structure is intrinsic to both the predictors (exposures) and to unmeasured background processes. An early formal recognition of this issue can be found in Clayton et al. 1993 , which demonstrated how including spatially varying error terms in the model changes covariate effect estimates, but this change was favored over the unadjusted estimate which is biased under confounding. Subsequent works have had differing opinions about this ( Woodard et al. 1999 , Reich et al. 2006 , Wakefield 2007 , Hodges & Reich 2010 , Schnell & Papadogeorgou 2020 , Khan & Calder 2022 , Zimmerman & Ver Hoef 2022 ) with arguments put forth both for and against including spatially correlated errors in the model. A particularly influential contribution by Paciorek 2010 quantified this bias in generalized least squares (GLS) settings, relating it to the relative smoothness of the exposure and the unmeasured confounder. This work also underscored the identifiability challenges of separating the exposure effect from the influence of spatial confounding, unless certain restrictions or structural assumptions are imposed. Extensive theoretical and empirical results in Khan & Berrett 2023 largely agree with the conclusions of Paciorek 2010 . Related work in spatial prediction frameworks, such as that of Page et al. 2017 , arrives at similar conclusions. Gilbert et al. 2025 proved that the GLS estimator is consistent under varied forms of spatial confounding as long as the exposure has some non-spatial variation. Dupont et al. 2022 proves consistency of their spatial+ method under a similar assumption of Gaussian noise in the exposure. Gilbert et al. 2021 grounded the spatial confounding problem in the framework of causal inference, going beyond the linear regression case. Their sufficient conditions for identifiability of a causal effect under confounding also assumed some non-spatial variation in the exposure. Things get more complicated if the exposure is only a function of space with no non-spatial component. Such examples are common in many environmental studies, e.g., studying the effect of distance to fracking mines on health ( Casey et al. 2016 , Currie et al. 2017 ) or of distances to roadways on air-quality ( Karner et al. 2010 ) . Brugge et al. 2013 explicitly raises concerns about the inability to adjust for co-smooth spatial confounders. Estimators like GLS and spatial+ whose consistency is guaranteed in the presence of noise in the exposure, may no longer work in this smooth-exposure setting. In fact, it has been proven that when exposures are too smooth relative to the confounder, the exposure effect cannot be identified even in a linear regression setting ( Bolin & Wallin 2024 , Datta & Stein 2025 ) . In the setting of smooth exposures, most methods rely on the assumption that confounders vary mainly at broad spatial scales, while exposures may display more localized, high-frequency spatial patterns. Extensive experiments Paciorek 2010 and Khan & Berrett 2023 have shown that such a situation is favorable for recovery of the exposure effect as the exposure becomes unconfounded at higher frequencies. This, along with the assumption of a linear relationship between the outcome and the exposure, common in spatial setup, solves the problem of spatial confounding. Methods that build on this idea include both parametric and semi-parametric approaches, such as those proposed in Thaden & Kneib 2018 , Keller & Szpiro 2020 , Dupont et al. 2022 , and Guan et al. 2023 . When the unmeasured confounder has the same or finer spatial variation than the exposure, existing techniques generally fail, and standard spatial estimators can perform worse than even the unadjusted estimator as seen in Paciorek 2010 , Khan & Berrett 2023 . This can be viewed through the lens of causal inference as a violation of the positivity assumption , which requires that all levels of exposure be observed across levels of confounders. Nonetheless, there are important cases where this assumption does not hold. As mentioned earlier, spatial contours of exposures like distances to industrial sites, greenery, or major roadways can be quite smooth, precluding the use of techniques relying on presence of high-frequency variation unique to the exposure. In fact, the notion of smoothness defined by the ordering of basis functions may not generalize to two or higher-dimensional spatial domains. For example, using kronecker product of basis functions of one-dimensional basis, creates a two-dimensional basis with two indices dictating smoothness (one along each coordinate). Approaches like those of Keller & Szpiro 2020 which model the confounder using basis functions of low-indices do not easily generalize for such basis sets. In this work, we develop a general framework for effect estimation in the spatial linear regression model that adjusts for unmeasured spatial confounders, without assuming that the confounders are smoother than the exposure. Instead, we introduce a milder plurality rule assumption, inspired by ideas in the robust instrumental variable (IV) literature. The central assumption is that, in a representation of the exposure X and the confounder U along a set of orthonormal basis functions, the set of ratios of the projections of U and X on each of the basis functions has a larger cluster around zero than around any other value. This assumption subsumes the rougher exposure assumption as a special case. Under plurality,