Sensitivity Analysis Risk Characterization Method
106 methods ANOVA and variance-based method Mokhtari and Frey 2005; Ellouze et al. 2010
were tested and gave similar results as the one presented in this report. 8.4.1.
Changing one Factor at a Time
One way to study the model is to evaluate the change in the output following the change in one input. In order to study the model, some artificial scenarios will be tested to evaluate their impact
on the risk per serving for a specific country and a particular subpopulation say: Canada, Elderly. The tested scenarios are:
- Environmental contamination: from pasteurized-milk cheese baseline contamination level
section 6.4, Table 20-Table 21 to exactly 1, 10, 100, 1,000, 10,000, 100,000 cfucheese; and from pasteurized-milk cheese baseline prevalence section 6.4, Table 18 to
prevalences less than the 20 point of that distribution; -
Growth characteristics: from pasteurized-milk cheese baseline section 6.1, Table 12- Table 13 to EGR
20
s equal to 0, ½ baseline, 2 × baseline; from baseline section 6.1, Table 14 to maximum population densities equal to the baseline -1 log
10
or to the baseline +1 log
10
; -
Transport and Marketing temperature: pasteurized-milk cheese baseline -1°C, pasteurized-milk cheese baseline +1°C;
- Temperature at Retail: pasteurized-milk cheese baseline -1°C, pasteurized-milk cheese
baseline +1°C; -
Home storage temperature: pasteurized-milk cheese baseline -1°C, pasteurized-milk cheese baseline +1°C;
- Home storage duration: maximum duration of storage at home of 28 days compared to 56
days.
8.4.2. Rank Correlation
The second method used is an evaluation of the Spearman’s rank correlation between inputs to the model and outputs of the model Frey and Patil 2002. This method’s output is frequently
displayed as a “tornado chart.” While frequently used in risk assessment, this sensitivity analysis also remains rough Borgonovo 2006. Exploring interactions that could occur in the model is
107 difficult and the method is insensitive to several important types of dependence between output
and inputs: non-linearity and thresholds. Considering both the variability and uncertainty, the following impacts of parameters on the final
outputs were explored: i impact of variable parameters on outputs: for a specific input-output pair, one
Spearman’s rank correlation may be estimated for each of the N
u
simulations, leading to N
u
Spearman’s rank correlations. The median, the 0.025
th
quantile, and the 0.975
th
quantile of these N
u
values may then be used as an estimate and credible interval of the Spearman’s rank correlation for that pair;
ii impact of uncertain parameters on outputs: in the uncertainty dimension, one can estimate the Spearman’s rank correlation between uncertain parameters N
u
values and some statistics evaluated in the variability dimension, such as the mean or specific
quantiles of the risk outputs’ variability.