Background
The Sensitivity Analysis study provides global sensitivity analysis and supports both the Sobol indices method and the correlation method.
For the Sobol indices method, the most efficient method is to generate a polynomial chaos expansion (PCE) surrogate model for each quantity of interest (QoI). The Sobol indices can readily be computed from the coefficients of the PCE model without sampling through a Monte Carlo procedure. The Sobol indices can also be computed through Monte Carlo analysis when a Gaussian process (GP) surrogate model is used. For details on the surrogate models and how the Sobol indices are computed, see the following sections in the Theory chapter: Surrogate Models — Polynomial Chaos Expansion, Surrogate Models — Gaussian Process, and Sensitivity Analysis — Sobol Index.
For the correlation method, the sensitivity analysis is performed through pure sampling-based statistical analysis with no surrogate model. There are four types of correlations:
•
The bivariate correlation, also known as the Pearson’s correlation, which computes the linear relationship between each input parameter and QoI.
•
The rank bivariate correlation, also known as Spearman’s correlation, which computes the monotonicity between the each input parameter and QoI.
•
The partial correlation, which computes the linear relationship between each input parameter and QoI where all linear effects from other input parameters are removed.
•
The rank partial correlation, which computes the monotonic relationship between each input parameter and QoI where all linear effects from other input parameters are removed. For details on how each correlation is computed, see Sensitivity Analysis — Correlations in the Theory chapter.