Step 5: Rank Uncertainties

The primary analysis objective is to quantify the uncertainty of the results of the model application. This objective was achieved in Step 4 of the analysis. The message may be that the uncertainty of the model results is small enough so as to be of no concern. If this is the case, the present analysis step may be skipped. The uncertainty may, however, be such that meaningful decision-making is impossible or that the subjective probability for the violation of a limit value is too high. In this case, it will be of vital interest to rank the uncertain parameters, model formulations and input data with respect to their contribution to model result uncertainty. This task is known as “uncertainty importance analysis”. The term “sensitivity analysis”, although frequently used, is misleading as will be shown below. Uncertainty importance measures from correlation, regression and variance decomposition, using raw or rank transformed data, are formally derived and are discussed with respect to their benefits and shortcomings. This discussion summarizes with a table of recommendations for the choice of importance measures in four problem categories. Section 6.4 presents a method for the identification of uncertainties that tend to be responsible for the upper (or lower) quantile values of the subjective probability distribution of a model result. Section 6.5 illustrates, by practical cases, how importance analysis helps to improve the reliability of computer models and to enhance the quality of their results. The graphical presentation of importance rankings is discussed in Sect. 6.6.