CONTAMINANTS IN SEDIMENTS - USING ROBUST REGRESSION FOR GRAIN-SIZE NORMALIZATION

Sediment contaminant concentrations usually show an inverse correlatio n with gain size. This can cause difficulties in distinguishing real d ifferences in contamination from artifacts caused by variations in sed iment texture. To overcome this, regression analysis is frequently use d to remove the dependency of concentrations on gain size. However, le ast squares regression lines can be affected markedly by the presence of a small number of unusual samples in the dataset. These outliers ma y represent samples which are more severely contaminated or which were derived from areas with different underlying geology. They can be rem oved semi-manually, but robust regression methods such as least absolu te values provide a convenient and objective alternative. The methods are illustrated using an example dataset of metal contaminants in sedi ments from the Humber Estuary, United Kingdom. Least squares regressio n on the complete dataset yields a rather poor gain size normalization for several elements. By contrast, least absolute values regression p roduces results very similar to those obtained by least squares regres sion after careful manual removal of outliers, but it avoids the need for subjective judgments of which data points to omit from the analysi s. The intercepts of several of the fitted regression lines were non-z ero, indicating that regression-based normalization is preferable to m ethods based on ratios.