Consistency of common spatial estimators under spatial confounding

This paper addresses the asymptotic performance of popular spatial regression estimators on the task of estimating the linear effect of an exposure on an outcome under "spatial confounding" -- the presence of an unmeasured spatially-structured variable influencing both the exposure and the outcome. The existing literature on spatial confounding is informal and inconsistent; this paper is an attempt to bring clarity through rigorous results on the asymptotic bias and consistency of estimators from popular spatial regression models. We consider two data generation processes: one where the confounder is a fixed function of space and one where it is a random function (i.e., a stochastic process on the spatial domain). We first show that the estimators from ordinary least squares (OLS) and restricted spatial regression are asymptotically biased under spatial confounding. We then prove a novel main result on the consistency of the generalized least squares (GLS) estimator using a Gaussian process (GP) covariance matrix in the presence of spatial confounding under in-fill (fixed domain) asymptotics. The result holds under very general conditions -- for any exposure with some non-spatial variation (noise), for any spatially continuous confounder, using any choice of Mat\'ern or square exponential Gaussian process covariance used to construct the GLS estimator, and without requiring Gaussianity of errors. Finally, we prove that spatial estimators from GLS, GP regression, and spline models that are consistent under confounding by a fixed function will also be consistent under confounding by a random function. We conclude that, contrary to much of the literature on spatial confounding, traditional spatial estimators are capable of estimating linear exposure effects under spatial confounding in the presence of some noise in the exposure. We support our theoretical arguments with simulation studies.