B. Jones writes: In an interesting paper Norry et al. 1994 described the geochemistry and mineralogy of the Peterborough Member of the Oxford Clay Formation from the UK Jurassic. A significant part of this work concerns the relationships between the 36 major, minor and trace elements determined. The inter-element relationships are described by correlation coefficients calculated between the Al2O3 normalized elemental data, and the ‘index variables’ SiO2/Al2O3, organic-C/Al2O3, and P2O5/Al2O3. Based on the magnitudes of these correlations the probable mineralogical residences of the elements are discussed, and elements classified as belonging either to a silica group, an organic carbon group, or a phosphate group. However, the methodology employed would result in significantly large correlations even if the initial elemental variables were random, and not correlated. As a consequence, geological interpretation of data analysed by this method may be unreliable. Having previously employed similar methodology, my purpose here is to prevent further workers following this same path without being aware of the pitfalls, and to stimulate further discussion regarding the statistical treatment of geochemical data.
The use of correlation in the interpretation of the relationships between elemental variables in geochemistry is dogged by the constant sum problem. This has been recognized by geochemists for many years. Ratios of the variables are free of the constant sum effect and Aitchison (1986) has suggested a new method for the analysis of relationships between the variables in compositional datasets which involves taking the natural logarithms of ratios of the original variables, and analysing the