Spatial autocorrelation: bane or bonus?

Abstract

Spatial autocorrelation is a general phenomenon within biogeographical studies. However, considerable confusion exists about how to analyse spatially autocorrelated data collected using classical sampling methods (e.g. simple random sampling). We show that the two common discordant views about autocorrelated data depend upon the desired scale of inference. Although inferential statistics seeking to generalise to different unsampled spatial areas need to be adjusted for autocorrelation, if the inference is restricted to the area from which samples have been taken, then standard tests are applicable. In the latter case, incorporating autocorrelation into the model may actually improve the precision and power of the analysis. We found that the scale of spatial inference is rarely discussed, despite being of central importance to any spatial analysis. We suggest that spatial inference should rarely be formally generalised to unsampled areas, since it typically requires some assumption of stationarity, and is thus vulnerable to accusations of \"pseudo-replication\".