I'm wondering to what extent (if any) subsampling of observations can be used to account for spatial autocorrelation within data.
Is taking a smaller sample (subsample) of observations (without replacement) a valid remedy to deal with spatial autocorrelation?
The reason I ask is because I have genetic DNA sequence data generated from many individuals for different animal species. Much of this data lack accompanying GPS coordinate data identifying where each individual was collected, which would be of immense use in saying something regarding population structure.
I will be employing univariate semiparametric regression, specifically Generalized Additive Models (GAMs), Shape-Constrained Additive Models (i.e., monotonic/concave GAMs, called SCAMs) and Kriging. I have a set of 10 candidate models (4 GAMs, 3 SCAMs and 3 Kriging models) and I am using the Akaike Information Criterion (AIC) to select the most parsimonious model that adequately explains variation in observed data with the 'mgcv' R package.
To extend on this, typically biologists sample individuals of a species across their geographic ranges. Take humans for example: distinct subgroups exist and these subpopulations correspond to ethic races. Different ethnic groups will have different predispositions to disease, which are often correlated (highly) with geography.
There exists much in the primary literature on "spatial subsampling", especially in regard to bootstrapping, but I've not been able to find anything to answer my general question. Looking to the CV site also did not yield an answer.
Any insight would be greatly appreciated and warmly welcomed.