Multivariate aspects of model uncertainty analysis: tools for sensitivity analysis and calibration

This paper deals with the analysis of complex environmental models. It is argued that such models are needed to predict environmental change outside current conditions. The use of these models poses some hard questions about their validity, and it is concluded that they should be termed useful rather than scientific in the strict sense. The complexity of the models arises from the high number of input-parameters and output variables. The paper proposes statistical and visualization techniques to explore the relations between the two. Results of a sensitivity analysis can be grouped in order to identify different modes of model behaviour. Within each mode an ordering of sensitivity coefficients is possible. In model calibration there is often a trade-off in the fit to various model output. Quantification of these relations can help in a proper choice of weights and serve as a tool in model analysis. A common feature of the techniques proposed in this paper is that they recognize modelling as a series of judgements to be made by the modeller. By providing more insight into this process these choices can be made more explicit, less arbitrary and more reproducible.