Sensitivity Analysis

C Pichery, EHESP (French School of Public Health), Rennes, France

Sensitivity Analysis: Definition and Properties The objectives of a best global SA are as follows:
l Cope with the influence of the scale and shape: identify and
In a numerical (or otherwise) model, the Sensitivity Analysis
prioritize the most influential inputs and determine the
(SA) is a method that measures how the impact of uncertainties
influential entries not to make them consistent by incor-
of one or more input variables can lead to uncertainties on the
porating the effect of the range of input variation and the
output variables. This analysis is useful because it improves the
form of its probability density function
prediction of the model, or reduces it by studying qualitatively
l Be able to treat grouped variables as if they were single
and/or quantitatively the model response to change in input
variables
variables, or by understanding the phenomenon studied by the
l Include multidimensional averaging
analysis of interactions between variables. However, the target
l Map the behavior of the output relative to inputs, focusing
of interest must not be the model output per se, but the
in specific areas if necessary
question that the model has been called to answer.
l Be model independent
In other words, the expected values of various parameters
l Calibrate the model variables with respect to certain infor-
involved can be used to evaluate the robustness, i.e., ‘sensitivity’
mation available (observations real output, constraints)
of the results from these changes and identify the values beyond
which the results change significantly. SA identifies priority There are three methods of global SA of a model.
needs for improving knowledge. Indeed, this analysis reduces The first method is the screening that can screen roughly the
the uncertainties of the parameters of the assessment and then, input variables most influential among many. Screening
decisions about the phenomenon under study can be taken. qualitatively analyzes the importance of input variables on the
variability of the response of the model. This approach helps to
establish a hierarchy within the input variables according to
Two Methods of SA their influence on the variability of the response and to identify
the subset of factors that control most of the output variability
Local SA
with a relatively small effort of calculation. Typical screening
The local analysis differs from the global analysis by looking at the designs are one-at-a-time with the method of Morris, in which
value of the response; in fact, it studied how small perturbations the impact of changing the values of each of the chosen factors
around a value of the inputs affect the output value. This analysis is evaluated in turn.
therefore determines the local impact of changes in input The second method is that measures of importance: It helps
parameters, in a small interval around a nominal value of output prioritize accurately the influence of each input variable on the
parameters. This deterministic approach is estimating the indices model output. This approach generally uses Monte Carlo anal-
of sensitivity based on the partial derivatives of the model at ysis and several statistical tools that show quantitatively the
a specific point. Local SA is usually carried out by computing influence of each input variable. These tools are linear regression
partial derivatives of the output functions with respect to the with the following sensitivity main measures: Pearson coeffi-
input variables (differential analysis). For risk analysis, when the cient, standardized regression coefficients, partial correlation
objective of the local SA is to quantify the probability of exceeding coefficients, standardized rank regression coefficients, and
a threshold, the first-order reliability method (FORM) and the partial rank correlation coefficients. Statistical tests can supple-
second-order reliability method (SORM) measure the sensitiv- ment this regression (Fisher test and Kruskal–Wallis test). When
ities with respect to this threshold which is exceeded. However, the model studied is nonlinear and nonmonotonic, the
this local approach has limitations that are the assumptions of importance of the inputs on the output of the model is estimated
linearity and normality, and local variations, results are limited to using the variance decomposition function. The method of the
the neighborhood of a nominal point, and cross effects cannot be variance decomposition can also be performed by the method of
detected. Sobol or by a model simulation which runs based on efficient
sampling, using different techniques: high-dimensional model
representation technique, Fourier amplitude sensitivity test
Global SA
(FAST), and the extended FAST method.
The global SA interests the entire field of possible variation of The third method uses the tools of exploration of the model to
the input variables, by determining the input variables of measure the effects of inputs throughout their range of variation.
a model that contribute most to an amount of interest calcu-
lated using this model. In other words, this analysis determines
how much uncertainty about the response is due to the SA Applications
uncertainty of each input variable (or group of input variables),
i.e., how much of the output variance is due to such inputs or Note that many application domains are interested in the SA.
that set of inputs. In the field of risk assessment, SA can be applied to identify

236 Encyclopedia of Toxicology, Volume 4 http://dx.doi.org/10.1016/B978-0-12-386454-3.00431-0

both the variables that dominate risk estimates than those that Further Reading
are relatively small.
In quantitative toxicology, uncertainties can be reduced in Iooss, B., 2011. Review of global sensitivity analysis of numerical models. J. Soc. Fr.
Stat. 152, 3–25.
the formulation and application of the (integrated or not) Jacques, J., 2005. Contributions à l’analyse de sensibilité et à l’analyse discriminante
models by applying routinely the SA methods and by inte- généralisée. Thèse de l’Université Joseph Fourier, Grenoble.
grating them as part of model evaluation and application. Krishnan, K., Andersen, M.E., 2010. Quantitative Modeling in Toxicology. John Wiley &
Techniques used for physiologically based pharmacokinetic Sons, Chichester, UK.
Saltelli, A., 2004. Sensitivity Analysis in Practice: A Guide to Assessing Scientific
(PBPK) models need to be numerically efficient, able to deal
Models. John Wiley & Sons, Chichester, UK.
with nonmonotonic output and differential forms. Most of
these SA methods satisfy these criteria and allow for variance
decomposition and estimation of main and interaction effects. Relevant Website
These PBPK models have generally a large number of input
parameters (>20), and the global method is computationally http://sensitivity-analysis.jrc.ec.europa.eu/methods/ – EU Joint Research Centre:
expensive. One solution is to use a global screening method, Sensitivity Analysis.
such as the Morris test to determine the most influential
parameters and then to run a global quantitative method on
only those parameters.
SA can also be applied to nuclear engineering, chemistry,
ecology, medicine, and economics.

See also: Uncertainty Analysis; Risk Assessment; Uncertainty