A Bayesian multinomial Gaussian response model for organism-based environmental reconstruction

We present a Bayesian hierarchical multinomial regression model (Bummer) fo r organism-based quantitative paleoenvironmental reconstruction. The model is based on the classical (direct) approach to calibration and on careful s tatistical environmental modeling that takes account of statistical depende ncies among species. We compare our Bayesian model Bummer to seven other methods, including the widely used weighted averaging (WA) techniques and our previous Bayesian mo del Bum. The methods are evaluated on a surface-sediment chironomid trainin g set of 62 subarctic lakes in northern Fennoscandia by comparing the cross -validation prediction statistics of different models. Bummer outperformed other methods, yielding the smallest prediction error, the smallest bias, a nd the largest correlation coefficient. We conclude that the promising performance of our Bayesian multinomial Gaus sian response model is due to the following reasons: (i) the uncertainty co ncerning site specific latent variables is taken into consideration; (ii) e cological background knowledge is embedded to the model description; (iii) the species compositions are considered as a whole; and (iv) reconstruction is based on the classical approach to calibration.