I am trying to weigh up if kriging (or cokriging) the residual error from my model is a valid approach. My first step was to compute whether the error is spatially autocorrelated. I have done this using Moran's I statistic but I'm having trouble interpreting the results. Below are 5 examples (my actual data is the ERROR example), the text is the output from the ape
package in R
.
I interpret that as the p-value of RANDOM is >0.05 then the data is not autocorrelated. Reading the help file with ape
it states that if the observed values are significantly greater than the expected value then your data is positively spatially autocorrelated (e.g. E vs W and X coord). I would have expected the observed value for CHESS to be negative. How should I interpret the results from ERROR? The p-value is <0.05 so this would suggest some degree of spatial autocorrelation, but the difference between observed and expected values is relatively small compared to the other examples. Below is the (massive!) semivariogram for the ERROR example.
This again would suggest a degree of spatial autocorrelation. My question would then be would kriging this error be an acceptable approach to reduce error in my final output?