Mitigating the risk of bias exacerbation when controlling for unmeasured spatial confounding for binary exposures.

In the presence of unmeasured spatial confounding, spatial models can increase bias of exposure effect estimates compared to nonspatial models (bias exacerbation), leading to debate regarding their use. For continuous exposures, this issue predominantly occurs when exposures are purely spatially varying (eg, air pollution), whereas spatial models can mitigate bias for individual-level factors with nonspatial variability (eg, blood pressure). However, in the common setting of binary exposures, the potential for bias exacerbation and the optimal choice of spatial method have not been well characterized. Focusing on point-referenced data, through simulation and a real-world application, we compared approaches for controlling unmeasured spatial confounding for binary exposures, including spatial spline models, matching, and combinations of matching and penalized spline (PS) models (all targeting average treatment effect). Additionally, we generalized the exposure-PS (E-PS) model to accommodate binary exposures, which selects degrees of smoothing to explain variability in the exposure. Like continuous exposures, spatial methods were able to eliminate bias for individual-level binary exposures. Unlike continuous exposures, bias was also reduced for purely spatial binary exposures, provided spatial spline adjustment was sufficiently flexible. Generalized E-PS and the combination of nearest neighbor matching with replacement and PS generally resulted in the largest bias reductions.