Spatial Self-Confounding: Smoothness-related estimation bias in spatial regression models

About

Spatial regression models are widely used to capture the relationship between observations and covariates, employing Gaussian random fields to account for spatial variability not explained by the covariates. A new study by researchers David Bolin and Jonas Wallin addresses a critical yet often overlooked problem in these models: smoothness-related spatial self-confounding. The work examines how misspecified covariates, particularly when there are differences in smoothness between variables, can lead to severe and counter-intuitive biases in the estimation of regression parameters. These findings have profound implications for various scientific domains that rely on spatially referenced data, from environmental sciences to epidemiology.

The researchers studied the theoretical properties of the generalized least-square estimator and demonstrate that when covariates are "too rough" compared to the observed data, the estimated regression coefficients can converge to zero as the number of observations increases, even when there is high correlation between observations and covariates. Conversely, under certain conditions, the estimates can diverge to infinity. Both scenarios can lead to incorrect scientific conclusions, with the importance of rough covariates being severely underestimated or overestimated.

The authors demonstrated that these counter-intuitive behaviors are not merely theoretical curiosities but occur in real data. Through applications to temperature and precipitation data, the study shows both phenomena in practice. To address this problem, the researchers propose adding a smoothing step in the regression process and demonstrate, both theoretically and empirically, that this approach can correct the estimation bias. This solution offers practitioners a concrete tool to avoid misleading conclusions in spatial analyses, ensuring that the importance of covariates is not artificially diminished or exaggerated due to differences in smoothness. The work represents an important advance in understanding the limitations of spatial regression models and provides practical guidance for improving the reliability of statistical inference in spatial applications.

REFERENCE:

Bolin, D., & Wallin, J. (2025). Spatial self-confounding: Smoothness-related estimation bias in spatial regression models. Biometrika, in press.