Sensitivity analysis in the context of uncertainty analysis for computationally intensive models Original Research Article

Sensitivity analysis is often performed in connection with uncertainty analysis, the latter being the prime objective. In this context, the aim of sensitivity analysis is to identify the main contributors to model output uncertainty. Therefore the term “Uncertainty Importance Analysis” is sometimes used. In general, uncertainty analysis proceeds by Monte Carlo simulation. The affordable number of model runs is usually small for processor-time or computationally intensive models. This has consequences for the type of uncertainty statements needed. For efficiency reasons, sensitivity analysis cannot afford a separate specifically chosen set of model runs but has to use those that were performed for the purpose of uncertainty analysis. Since their number n is small and the number m of uncertainties is frequently large, correlation coefficients and standardized regression coefficients, the latter possibly obtained from stepwise regression, are a reasonable choice of sensitivity measures. Spurious correlation is inevitably present in the multivariate sample of size n and can often not be reduced or eliminated. The correlation ratio is an indispensable sensitivity measure, whenever model uncertainty is expressed by more than two model alternatives or when measures, quantifying degrees of linear or monotone relationship, are not adequate. As a consequence of the small sample size, the correlation ratio is affordable only in approximate form. Results from analyses of applications of computationally intensive models serve as examples. The last section presents a practical example intended to illustrate the outstanding role played by uncertainty and sensitivity analysis in the quality assurance of the computer model and of its application.