Premium
APPLYING BAYESIAN STATISTICS TO ORGANISM‐BASED ENVIRONMENTAL RECONSTRUCTION
Author(s) -
Toivonen Hannu T. T.,
Mannila Heikki,
Korhola Atte,
Olander Heikki
Publication year - 2001
Publication title -
ecological applications
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.864
H-Index - 213
eISSN - 1939-5582
pISSN - 1051-0761
DOI - 10.1890/1051-0761(2001)011[0618:abstob]2.0.co;2
Subject(s) - bayesian probability , bayesian inference , computer science , bayesian statistics , calibration , statistics , machine learning , inference , ecology , artificial intelligence , mathematics , biology
Reconstruction of environmental conditions from biological indicators, such as chironomids, can provide valuable insight into the variance of natural systems and the extent of current environmental problems. Various numerical techniques have been developed for this challenging task, based on different models and assumptions, but ignoring some of the uncertainties inherent in the reconstruction. We study the use of Bayesian modeling and inference in organism‐based palaeoenvironmental reconstruction. We propose a Bayesian model, BUM, and compare it empirically with eight other methods, including the widely used weighted averaging (WA) technique. The methods are evaluated on a surface‐sediment chironomid training data set from 53 subarctic lakes in northern Fennoscandia by comparing the prediction statistics of these data. The resulting calibration models are also applied to fossil chironomid assemblages in order to evaluate the differences in Holocene temperature reconstructions. The empirical results indicate that BUM is competitive compared with the state‐of‐the‐art methods. We also describe a generic Bayesian framework for reconstruction, to demonstrate Bayesian tools for reasoning with a variety of ecological response models. Bayesian statistics support the “classical” approach to regression and calibration whereas the state‐of‐the‐art methods, including WA, are based on the conceptually more controversial “inverse” approach. Further, the use of probability distributions rather than point estimates for species responses gives a principled method for handling uncertainty in response modeling.