|
Abstract:
|
Historically, when residuals in regression models were spatially correlated, it was common practice to include a spatial random effect. This practice was thought to reduce bias in the estimation of other covariate effects. Recently, research into a phenomenon known as spatial confounding has challenged this practice. Spatial confounding is often described as occurring when there is multicollinearity between a spatially smooth covariate and a spatial random effect. It is thought to lead to poor associational and causal inferential outcomes, and research into methods to mitigate it is ongoing. In this talk, we will explore how at least two distinct phenomena have become conflated with the term spatial confounding. Within the class of spatial linear mixed models, we show how multicollinearity can arise either in a data generation setting or a model fitting setting. While both issues are referred to as spatial confounding, they can differentially impact inference. Using illustrative examples, we examine both settings when the inferential goals are associational or causal.
|
