I'm looking to model land cover change using a variety of environmental predictors (e.g. elevation, rainfall, etc.) stored as raster layers. In most similar studies I've found in the literature the authors sample the raster layers at some subset of the cells before performing the regression. I'm trying to understand if this sampling is done purely for efficiency, i.e. because it would be prohibitively slow to run the analysis over the whole grid? Or, is there a statistical reason that it is better to use a subset of the data, e.g. to reduce spatial autocorrelation?