Résumé
Sensitivity analysis is frequently used to select the most influent parameters to be estimated from scarce available data. However, the capability of this approach to improve model predictions is not well known, especially for complex environmental models. This paper investigates the relevance of estimating the most influent parameters only and setting the other parameters to their nominal values. More precisely, an empirical relationship is established between the global sensitivity index of a parameter and the Mean Square Error of Prediction, for a dynamic model simulating greenhouse gas emission. The results show that the estimation of parameters with low sensitivity indices is likely to give poor model predictions whereas the estimation of the parameters with high indices leads systematically to a reduction of the mean square error of prediction.