Primary Submission Category: Causal Inference in Networks
Harnessing the Scale of Spatial Confounding for Causal Inference with Areal Data
Authors: Sophie Woodward, Mauricio Tec,
Presenting Author: Sophie Woodward*
Studies investigating the causal effects of continuous exposures on human health – such as air pollution, green space, or crime – often rely on observational and spatially-indexed areal data. A prevalent challenge is unmeasured spatial confounding, where an unobserved, spatially-varying variable affects both exposure and outcome, leading to biased causal estimates and invalid confidence intervals. Confounding might occur only at some spatial scales but not others. For instance, in studies on the health impacts of air pollution, local variations in healthcare may be less correlated with air pollution exposure after accounting for socioeconomic status and demographics, thus decreasing unmeasured confounding from healthcare at the local scale. Conversely, there are applications in which confounding is stronger locally but dissipates at coarser scales. We introduce a methodology to address spatial confounding bias restricted to certain spatial scales. By decomposing exposure into a component influenced by the unmeasured confounder and an independent component, based on assumptions about the scale of confounding, we can achieve causal identification of the exposure-response function by adjusting for the confounded component. This approach unifies existing literature where previous methods are special cases of our identifying functional. We also develop a sensitivity analysis framework that produces a sequence of estimators for different scale of confounding assumptions.