Skip to content
Prev 28300 / 29559 Next

eigenvectors increase spatial autocorrelation in ols regression

On Wed, 22 Jul 2020, Peter B. Pearman wrote:

            
You are using point support (your forests are at points not areas, so that 
the observations are not contiguous). The advice to consider this as a 
mixed model problem may be relevant. It is certainly the case that 
ncf::correlog() should not be used for regression residuals.

Further, your data are in a band 1.5-4.3E, 43.0-43.4N, so using 0.7 
degrees in dnearneigh() was a risky choice, and squashes the data. If you 
coerce to an sf object and apply an appropriate CRS, you get many fewer 
neighbours on average.

library(sf)
data <- st_as_sf(data, coords=c("LONG", "LAT"))
(nbnear4 <- dnearneigh(data, 0, 0.7))
st_crs(data) <- 4326
(nbnear4a <- dnearneigh(data, 0, 27))
# or project to a relevant planar UTM 31N spec.
data_utm31 <- st_transform(data, st_crs(32631))
(nbd <- dnearneigh(data_utm31, 0, 27))

and so on. SpatialFiltering() and ME() test against global residual 
autocorrelation in finding candidate eigenvectors, and running 
lm.morantest() on the fitted model shows that there is no global 
autocorrelation as intended (because positive and negative local 
autocorrelation cancels out). I think that the underlying problems are 
mixing approaches that are not asking the same questions, and in too 
little care in choosing the weights.

Hope this clarifies,

Roger