Univariate Patterns in the Design of Multivariate Analysis Techniques for Geochemical Data Evaluation

Abstract

The evaluation of regional geochemical data normally begins with a univariate analysis of each element. Patterns within the data are frequently recognized, but in cases where the patterns are obscure or not immediately understood, or where a large number of elements make interpretation difficult, a multivariate analysis is commonly undertaken. Unfortunately, this analysis is often made without consideration of the univariate results. Ignoring discernable univariate patterns in the design of subsequent multivariate analyses may lead to unneccessary ambiguities and/or complexities in the multivariate results and the subsequent interpretation. Geochemical concentration data from a soil survey in northern British Columbia are used to demonstrate this point. Univariate examination of the data reveals that most elements are bimodally distributed. However, the elements could be divided into two groups differing in the relative contribution of the two component populations (modes) making up their distributions. One group of elements has bimodal distributions with component populations of approximately equal size (55% and 45%); the other has bimodal distributions with component populations of vastly different size (90% and 10%).

This information was used in three subsequent principal component (PC) analyses. The PC loadings and scores from (1) an analysis of all elements in the data set (the standard approach) were compared with the loadings and scores from two other PC analyses. These include (2) an analysis which used only those elements with distributions containing component populations of approximately equal size (55:45), and (3) an analysis which used those elements with distributions containing component populations of vastly unequal size (90:10). Combined results from the two “structured” PC analyses (2 and 3) relate better to the geology and mineral deposits in the survey area than results from the wholesale, “unstructured” PC analysis (1). Furthermore, patterns from the “structured” PC analyses are more distinct than the wholesale PC analysis and lead to a less ambiguous and more complete interpretation.

The “structured” approach to multivariate analysis, which uses information from the univariate analysis in its design, thus represents a logical and superior data evaluation philosophy.