A statistical basis for multiobjective calibration

Abstract

Mathematical models are very powerful tools for improving the understanding of environmental systems. However, the output of such models usually deviates systematically from observations, and this bias is typically larger than the measurement error. These problems are mainly caused by the fact that the model is based on a simplified description of systems mechanisms that does not fully capture the real complex environment. Such errors lead to incorrect or biased estimates of parameters, model predictions, and uncertainties associated with these results. Furthermore, there is no objective way to decide how much bias to accept in each of several model outputs. A statistical (Bayesian) description of bias can address the former problem, and scientists use specialized calibration techniques to find good but unquantified compromises between bias in model outputs (multiobjective calibration). However, there is no straightforward way to estimate uncertainties in parameters and model predictions in such calibration techniques.