Community Research and Development Information Service - CORDIS

FP5

PACLIVA Report Summary

Project ID: EVK2-CT-2002-00143
Funded under: FP5-EESD
Country: Norway

Increased robustness of marine transfer functions

Palaeoenvironmental conditions can be reconstructed from microfossil assemblages with a range of numerical techniques, known as transfer functions, which use information on the modern relationship between species distributions and the environment. Many taxonomic groups have been used to reconstruct a wide variety of environmental variables. The selection of the best transfer function method to use, and the optimal values of metaparameters, is usually guided by the prediction error of the training set under cross-validation. Work done in PACLIVA has shown that all of the numerical methods used assume that the observations in the training set are independent, but that oceanic environments are highly autocorrelated.

This lack of independence between observations can severely bias the estimation of the true prediction error, and hence misinform model choice. The methods most sensitive to autocorrelation are those that seek local relationships between species and the environment (such as the modern analogue technique, and artificial neural networks). Because of their sensitivity, they give much lower uncertainty estimates than more robust methods (such as weighted averaging based methods), and are the most commonly used methods in marine palaeoecology.

This work suggests that some of the most widely used transfer function methods should only be used with caution, and some popular environmental proxies (such as dinoflagellate cysts) are much less valuable that previously thought.

Reported by

Bjerknes Centre for Climate Change
Allegaten 55
5007 Bergen
Norway
See on map
Follow us on: RSS Facebook Twitter YouTube Managed by the EU Publications Office Top