B0568
Title: A unified Bayesian approach to overcome spatial confounding in point-referenced data
Authors: Bora Jin - Johns Hopkins University (United States) [presenting]
David Dunson - Duke University (United States)
Abstract: Spatial confounding occurs when there is multicollinearity between covariates and spatial random effects. When spatial random effects correlated with a covariate are introduced into a model, estimation of the covariate effects is likely affected in terms of bias and uncertainty as the covariate and the spatial random effects compete for the same spatial signal in the response. To alleviate spatial confounding, a Bayesian approach is proposed, whose primary objective is accurate estimation of the fixed effect in terms of bias and coverage in point-referenced data, retaining all spatial signals in the data. The approach can accommodate any multicollinearity between covariates and spatial random effects and produce two kinds of covariate effects that separately appear in previous literature by decomposing the spatial random effects into the correlated part and the independent part with the covariates. The unified approach that simultaneously derives two kinds of fixed effects can lead to more robust inferences on the fixed effect than estimating each kind only.