Presentation: Spatial Confounding and Disease Mapping: Approaches for Inference on Smooth Regressors
Speaker: Ephraim Hanks, Ph.D., Assistant Professor of Statistics, Pennsylvania State University
Abstract: Spatially-referenced disease prevalence data can provide important clues to environmental risk factors, but recent work on spatial confounding has found that spatially-smooth covariates are often collinear with spatial random effects. We explain the mathematical nature of this confounding, and illustrate it using data on malaria from The Gambia. Some authors have suggested removing this confounding by constraining the spatial random effect to be orthogonal to the fixed effects, but we show that this approach suffers from increased Type-1 error rate under model misspecification. We suggest novel approaches to making inference on spatial data with spatially-smooth covariates that alleviate confounding in some situations. The first is to consider Laplace random fields, which provide more local spatial structure than their more common cousin, Gaussian random fields. The second is to consider causal inference based on Bayesian graphical models. The third is to consider more mechanistic models for spatial structure, when such information exists. We illustrate all approaches and compare and contrast their usefulness.