The Impact of Spatial Autocorrelations and Error Structures in Spatial Modeling
If you ﬁnd you still have signiﬁcant spatial autocorrelation in your residuals after including all the variables in your model that you and your theoretical framework think you should include, you cannot use the results from your model; these results will be biased, and in a predictable way. Speciﬁcally, in the presence of unaccounted for spatial autocorrelation, regression results are biased towards false positives, that is your results will be more likely to tell you a variable is a signiﬁcant predictor or explanatory factor in understanding your dependent variable than is really the case. The false positive problem is the most dangerous problem in statistical modeling, because you would conclude something is true when it is not. Suppose you are doing analyses for a city and planning and zoning decisions are inﬂuenced by the results? Suppose you are doing policy-related research on health outcomes and legal regulation of products and behaviors is at stake? Your results might suggest a product should be banned from the public because of its harmful eﬀects, when in fact the spatial nature of your data produced a false ﬁnding suggesting that this was the case.