ABSTRACT
In statistical inference, it is often necessary to exercise some care in applying a method that sounds like it is doing what one intends because the method may be doing something a bit different. Procedures used with ordinal correlations provide a case in point. Most texts provide 'tests of significance' for ordinal correlations that take the form of critical values that must be obtained in small samples at a given alpha level, and standard errors that can be used, along with assumed sampling normality for the estimates, in large ones. Zero correlation can occur in populations where there is some degree of curvilinear dependence between the two variables, and very often it seems implausible that two empirical variables will be completely independent. The consequences of ignoring non-normality in making inferences about correlations are unknown in any application, but the lack of robustness of r is substantial.
