I am attempting to test for spatial-autocorrelation among regression residuals via Moran's I in R using the 'spdep' package. While creating the spatial weights, changing the value of k changes the result of my Moran's I test. I have 1582 regions, and running the following code:
>Neighbour list object: Number of regions: 1582 Number of nonzero links: 158200 Percentage nonzero weights: 6.321113 Average number of links: 100 Non-symmetric neighbours list
I came across a semi-recent article: https://cran.r-project.org/web/packages/spdep/vignettes/nb.pdf that uses 8.717% non-zero weights, and I'm wondering if there is a best practice for finding the sweet spot for this type of procedure.