90
5 Handling Uncertainty and Sensitivity …
5.2.2 Stochastic Testing
The sub-models within ERA Acute are deterministic. To perform stochastic sensitivity testing, these models were made stochastic by using repeated random sampling
(Monte Carlo (MC)) methods (Marino et al. 2008): Instead of changing the values
one by one as in deterministic testing, they are assigned to a (a priori assumed) probability distribution. Configurations of model input values are then drawn randomly
from the probability distribution, and the resulting set of model outputs can be seen
as a random sample of the distribution of the output of interest (Helton et al. 2006).
Note that stochastic analyses are sensitive to the choice of probability distribution
used (e.g. Marino et al. 2008).
The result is a matrix with n values for each input parameter with corresponding
values for the model output (model predictions, results or endpoint) (Fig. 5.1). This
matrix is the input to the uncertainty and sensitivity analysis, which is performed
directly on the matrix. The sensitivity analyses were carried out using the Sampling
and Sensitivity Analysis Tool for Computational Modelling (SaSat) (Hoare et al.
2008a, b). For the sensitivity analysis, Pearson and Spearman correlation coefficient,
Partial Rank Correlation Coefficient (PRCC) analysis and Factor Prioritization by
Reduction of Variance (FPRV) were carried out (see e.g. Saltelli et al. 2000; Marino
et al. 2008). Combined, these methods can rank and quantify the most important
Fig. 5.1 Illustration of stochastic uncertainty and sensitivity analyses. The ERA Acute model
calculations are performed in the blue box. The uncertainty analyses are performed in Excel and
sensitivity analyses are performed with the MATLAB toolbox sampling and sensitivity analysis
tool for computational modelling (SaSAT)
5 Handling Uncertainty and Sensitivity …
5.2.2 Stochastic Testing
The sub-models within ERA Acute are deterministic. To perform stochastic sensitivity testing, these models were made stochastic by using repeated random sampling
(Monte Carlo (MC)) methods (Marino et al. 2008): Instead of changing the values
one by one as in deterministic testing, they are assigned to a (a priori assumed) probability distribution. Configurations of model input values are then drawn randomly
from the probability distribution, and the resulting set of model outputs can be seen
as a random sample of the distribution of the output of interest (Helton et al. 2006).
Note that stochastic analyses are sensitive to the choice of probability distribution
used (e.g. Marino et al. 2008).
The result is a matrix with n values for each input parameter with corresponding
values for the model output (model predictions, results or endpoint) (Fig. 5.1). This
matrix is the input to the uncertainty and sensitivity analysis, which is performed
directly on the matrix. The sensitivity analyses were carried out using the Sampling
and Sensitivity Analysis Tool for Computational Modelling (SaSat) (Hoare et al.
2008a, b). For the sensitivity analysis, Pearson and Spearman correlation coefficient,
Partial Rank Correlation Coefficient (PRCC) analysis and Factor Prioritization by
Reduction of Variance (FPRV) were carried out (see e.g. Saltelli et al. 2000; Marino
et al. 2008). Combined, these methods can rank and quantify the most important
Fig. 5.1 Illustration of stochastic uncertainty and sensitivity analyses. The ERA Acute model
calculations are performed in the blue box. The uncertainty analyses are performed in Excel and
sensitivity analyses are performed with the MATLAB toolbox sampling and sensitivity analysis
tool for computational modelling (SaSAT)