Geographical Gaussian Process GAMs: a better spatially varying coefficient model than multiscale GWR?

This paper proposes a novel spatially varying coefficient model for spatial regression using General Additive Models (GAMs) with Gaussian Process (GP) splines parameterised with observation locations. The brand leader in this area is probably Multiscale GWR (MGWR) models but these have a number of theoretical and technical limitations. Here, a GAM with GP spline model and a MGWR model were applied to simulated spatial datasets with varying degrees of spatial autocorrelation. The GAM was shown to perform better than MGWR under a range of fit metrics. Some unresolved issues are discussed such as model calibration or tuning of knots and spline smoothing parameters.