Spatial Confounding in GEEs
Date: 19 June 2020, Friday
Time: 4pm
Speaker: Dr Francis Hui (ANU)
Abstract:
Generalized Estimating Equations (GEEs) are a popular tool in many scientific disciplines for investigating the effects of covariates on the mean of a response. In the context of spatial analysis, GEEs rely on specifying a regression model for the marginal mean, a variance function, and a working correlation matrix characterizing the spatial correlation between observations. One of the key features of GEEs is that estimation of the covariate effects is robust to misspecification of the (spatial) working correlation matrix. That is, the choice of working correlation only affects efficiency and not the consistency (effectively, the target) of the GEE estimator.
In this talk, we introduce and explore the concept of spatial confounding in GEEs. Specifically, we show that in settings where the covariates included in the GEE are (also) spatially correlated, the choice of working correlation can in fact change the target coefficients one is estimating. Effectively, this arises due to an implicit multicollinearity occurring between the spatially correlated covariates and (the spatial effect induced by) the working correlation matrix. We propose the idea of a “restricted spatial working correlation matrix”, which estimates a so-called unpartitioned effect that pulls all the variability in the direction of the covariates into the marginal mean, and argue that is perhaps more in line with the underlying aim of GEEs. If time permits, we will touch upon the issues of standard error estimation as well as large sample properties.
Link: https://au.bbcollab.com/guest/fcf219c74ac743e89565a9e6e8d349a9
Video:
Date: 19 June 2020, Friday
Time: 4pm
Speaker: Dr Francis Hui (ANU)
Abstract:
Generalized Estimating Equations (GEEs) are a popular tool in many scientific disciplines for investigating the effects of covariates on the mean of a response. In the context of spatial analysis, GEEs rely on specifying a regression model for the marginal mean, a variance function, and a working correlation matrix characterizing the spatial correlation between observations. One of the key features of GEEs is that estimation of the covariate effects is robust to misspecification of the (spatial) working correlation matrix. That is, the choice of working correlation only affects efficiency and not the consistency (effectively, the target) of the GEE estimator.
In this talk, we introduce and explore the concept of spatial confounding in GEEs. Specifically, we show that in settings where the covariates included in the GEE are (also) spatially correlated, the choice of working correlation can in fact change the target coefficients one is estimating. Effectively, this arises due to an implicit multicollinearity occurring between the spatially correlated covariates and (the spatial effect induced by) the working correlation matrix. We propose the idea of a “restricted spatial working correlation matrix”, which estimates a so-called unpartitioned effect that pulls all the variability in the direction of the covariates into the marginal mean, and argue that is perhaps more in line with the underlying aim of GEEs. If time permits, we will touch upon the issues of standard error estimation as well as large sample properties.
Link: https://au.bbcollab.com/guest/fcf219c74ac743e89565a9e6e8d349a9
Video: