Computational Statistics & Data Analysis
Sensitivity analysis of model output: an investigation of new techniques
Computational Statistics & Data Analysis
Sensitivity analysis of model output: variance-based methods make the difference
Proceedings of the 29th conference on Winter simulation
Computational Statistics & Data Analysis
Computers and Electronics in Agriculture
Hi-index | 0.03 |
The present article is a sequel to an earlier study in this journal (Saltelli et al., 1993) where two new sensitivity analysis techniques were presented. Those techniques, the modified Hora and Iman importance measure (HIM^*) (Hora and Iman, 1986; Iman and Hora 1990; Ishigami and Homma, 1989, 1990) and the iterated fractional factorial design (IFFD) (Andres, 1987; Andres and Hajas, 1993) were proposed in order to overcome limitations in existing methods (Saltelli and Homma, 1992). Sensitivity analysis (SA) of model output investigates how the predictions of a model are related to its input parameters. In particular, Monte Carlo-based SA attempts to explain the uncertainly in model output by apportioning the total output uncertainty to the uncertainties of individual input parameters. It was pointed out in Saltelli and Homma (1992) that techniques employed in the existing literature were affected by severe limitations in the presence of nonmonotonic relationships between input and output. The search for better SA methods was pursued with reference to their ''reproducibility'' and ''accuracy''. The former is a measure of how well SA predictions are replicated when repeating the analysis on independent samples taken from the same input parameter space. The latter deals with the correctness of the SA results. The present note continues and completes the analysis of the performance of IFFD with respect to the two requirements. IFFD was found to generate highly reproducible results for sufficiently large sample sizes. It exceeded the capability of linear methods by detecting quadratic effects in the relationship between input parameters and model predictions, but had difficulty in dealing with higher order effects.