Robustness of spatial autocorrelation tests

Spatial autocorrelation expresses the dependence between values at neighbouring locations.
Testing if the spatial autocorrelation is significant may confirm or deny the need to
consider spatial analysis over the classical one. Indeed, the analysis of spatial data is only
meaningful if the spatial components bring information.
Several measures of spatial autocorrelation are defined in the literature. Moran’s index,
Geary’s ratio and Getis-Ord statistic are the most used statistics. Tests based on these
measures have been developed in the literature using asymptotic and permutation results.
They are used in practice in many fields, for instance in geography, economics, biogeosciences,
medicine, ... However, these tests should be cautiously applied because they are
not robust. A single contaminated observation can significantly modify their results.
This poster is composed of two parts. Firstly, the robustness of already available tools
for measuring spatial autocorrelation will be analysed. Secondly, alternative methods will
be proposed to robustly test the spatial autocorrelation.