Influence of the limit of detection on classification using otolith elemental signatures

Abstract

Otolith elemental composition is a commonly used tool in the classification of fish stocks. Lower detection limits associated with trace element analysis techniques are the source of much debate among fisheries biologists. Using empirical data from Atlantic herring (Clupea harengus) from two separate origins in the British Isles, we addressed methods of treating data at or falling below detection limits and the implications of each treatment for discrimination and classification analysis. Estimating actual values of non-detected concentrations, through imputation, is now a plausible and simply applied strategy and should be considered as an alternative to constant replacement. We compare a robust regression approach to estimating values for non-detects alongside a variety of constant replacement methods. We also show the value of a highly censored variable in understanding population discrimination; despite 40% of readings for one element being non-detects, we found that its removal was detrimental to the discriminatory power of the analysis.