Is 'Statistix Inferens' Still the Geographical Name for a Wild Goose?

After the recent death of Peter Gould, I decided to look once again at his insightful paper`Is Statistix Inferens the Geographical Name for a Wild Goose?' (Gould 1970) ± hence the title for this guest editorial. For those readers who have not seen this article, Gould outlines a number of shortcomings of the common statistical practices of geographers of the day. Many of these relate to the assumptions made in the statistical models used: • The functional form used in regression models (typically linear) • The random nature of the data sample • The probability distributions of random variates and error terms (typically normal) • The assumption of independence of random variates and error terms. He argues that often these assumptions are inappropriate in the geographical context, and goes on to question the validity of statistical inferences made using techniques based on these assumptions. In particular he finds the assumption of independence problematic in a spatial context. I found the paper to be as thought provoking now as it must have been when it was first published 30 years ago. One obvious question that arises is`to what extent have these issues been addressed since 1970?'. The answer is not a particularly simple one. The years between 1970 and now have seen the creation of a large number of new statistical techniques. Other techniques in their infancy in 1970 have now reached a stage of maturity. Many of these relax one or more of the above assumptions ± generalized additive modelling, nonparametric regression, kernel density estimation, randomisation tests and regression models with autocorrelated errors readily come to mind, and these by no means provide an exhaustive list. So why do I claim the answer to my earlier question is not simple? Firstly, one must ask to what extent these methods have diffused from the statistical community into the GIS literature. Although this has happened to some extent, there are still plenty of studies published which rely on non-spatial data analysis techniques such as least squares regression modelling. Why does the use of such methods continue? In some situations these methods may be genuinely appropriate ± this could only be the case if the relative or absolute geographical locations associated with the data played no part