Frey

1
IDENTIFICATION AND REVIEW OF SENSITIVITY ANALYSIS METHODS

By

H. Christopher Frey, Ph.D.
Civil Engineering Department
North Carolina State University
Raleigh, NC.

Sumeet R. Patil
Raleigh, NC.

Abbreviated Title: Sensitivity Analysis Methods

Correspondence:
H. Christopher Frey
311 Mann Hall, Campus Box 7908
Raleigh, NC 27695-7908.
Email: [email protected]

2
ABSTRACT

Identification and qualitative comparison of sensitivity analysis methods that have been
used across various disciplines, and that merit consideration for application to food safety risk
assessment models, are presented in this paper. Sensitivity analysis can help in identifying
critical control points, prioritizing additional data collection or research, and verifying and
validating a model. Ten sensitivity analysis methods, including four mathematical methods, five
statistical methods and one graphical method, are identified. The selected methods are compared
on the basis of their applicability to different types of models, computational issues such as
initial data requirement and complexity of their application, representation of the sensitivity, and
the specific uses of these methods. Applications of these methods are illustrated with examples
from various fields. No one method is clearly best for food safety risk models. In general, use of
two or more methods, preferably with dissimilar theoretical foundations, may be needed to
increase confidence in the ranking of key inputs.

Key Words: Sensitivity analysis methods, food safety, microbial risk assessment, critical
control points.

3
1.0 INTRODUCTION.
Concern for the safety of the food supply is motivated by recognition of the significant
impact of microbial food borne diseases in terms of human suffering and economic costs to the
society and industry, and an increasing global food trade (Lammerding, 1997). Mead et al.
(1999) have reported that food borne disease results in 76 million human illnesses in the United
States each year, including 325,000 hospitalizations and 5,200 deaths. ERS (2001) estimated the
cost of food borne disease to be $6.9 billion annually. Food safety is gaining increased attention
because of several trends. New food borne pathogens are emerging (Buchanan, 1996). Larger
batch production, distribution, and longer shelflife of food products contribute to broader
exposure to contaminating events. The proportion of food consumed that is supplied by food
services is growing and this reduces consumer control over food handling and processing.
Demand for safer food is growing as consumers are becoming more affluent, live longer, and are
better informed about diet (McKone, 1996). International trade in food also can introduce new
sources of risks to food safety, such as cattle or meat acquiring infection overseas.

Food safety regulatory agencies are taking a new approach to ensuring the safety of food
supply based upon the Hazard Analysis Critical Control Points (HACCP) system (Seward,
2000). One step in the HACCP system is to determine critical control points (CCP) where risk
management efforts can be focused. Food safety depends on many factors, such as composition
and preparation of the product, process hygiene, and storage and distribution conditions
(Zwietering and Gerwen, 2000). Given data gaps and the complexity and dynamic nature of the
food processing, transportation, storage, distribution, and preparation system, determining which
4
of the many nodes in the farm-to-table pathway constitute CCPs and represents a substantial
analytical challenge (Rose, 1993; Buchanan et al., 2000).

Sensitivity analysis of risk models can be used to identify the most significant exposure
or risk factors and aid in developing priorities for risk mitigation. For example, Baker et al.
(1999) identified sensitivity analysis as one of the principal quantitative techniques used for risk
management in the United Kingdom. Jones (2000) indicated that sensitivity analysis can provide
the basis for planning adaptation measures to mitigate the risk of climate change. Sensitivity
analysis can be used as an aid in identifying the important uncertainties for the purpose of
prioritizing additional data collection or research (Cullen and Frey, 1999). Furthermore,
sensitivity analysis can play an important role in model verification and validation throughout
the course of model development and refinement (e.g., Kleijnen, 1995; Kleijnen and Sargent,
2000; and Fraedrich and Goldberg, 2000). Sensitivity analysis also can be used to provide
insight into the robustness of model results when making decisions (e.g., Phillips et al., 2000;
Ward and Carpenter 1996; Limat et al., 2000; Manheim 1998; and Saltelli et al., 2000).
Sensitivity analysis methods have been applied in various fields including complex engineering
systems, economics, physics, social sciences, medical decision making, and others (e.g., Oh and
Yang, 2000; Baniotopoulos, 1991; Helton and Breeding, 1993; Cheng, 1991; Beck et al., 1997;
Agro et al., 1997; Kewley et al., 2000; Merz et al., 1992).
1.1 Objectives and Motivation.
The objectives of this paper are to identify, review, and evaluate sensitivity analysis
methods for applicability to risk assessment models typical of those used, or expected to be used,
5
in the near future for food safety risk assessment. As the key objective in this task is to benefit
from knowledge across many disciplines, the literature review extends well beyond food safety
to include many other fields. This objective serves as an aid in identifying potential CCPs along
the farm-to-table continuum, to inform decisions about food safety research and data acquisition
priorities, and to contribute to the development of sound food safety regulations.
1.2 Importance of Risk Assessment in Food Safety.
The objective of the risk assessment is to answer three risk questions (Kaplan and
Garrick, 1981):
What can go wrong?
How likely is that to happen?
What would be the consequences if it did go wrong?
Risk assessment provides structured information that allows decision makers to identify
interventions that can lead to public health improvement or to avoid future problems.

The HACCP system is used to prevent hazards associated with foods. The best
information of the risk from such hazards that industry may have today is qualitative, such as
whether a hazard presents a high, medium, or low risk (WHO, 2001). The outcome of a
HACCP-based system should be improved food safety assurance, measured by reduction in risk
to the consumer and not merely a reduction in the level of a hazard in a food (Hathaway, 1995).
Risk assessment can be used to determine which hazards are essential to control, to reduce, or to
eliminate (Buchanan, 1995; Hathaway, 1995; and Notermans et al., 1995). Therefore, risk
assessment can help in developing more effective HACCP plans.
6
Risk assessments can play an important role in international trade by ensuring that
countries establish food safety requirements that are scientifically sound and by providing a
means for determining whether different standards provide equivalent levels of public health
protection (WHO, 2001). Without a systematic risk assessment, countries may set requirements
that are not related to food safety and can create artificial barriers to trade.
1.3 Risk Assessment Framework.
As defined by the Codex Alimentarius Commission (CAC), risk assessment is a
scientifically based process consisting of four main steps: hazard identification, hazard
characterization, exposure assessment, and risk characterization (FAO, 1999). These steps are
similar to those defined in the risk analysis framework of the National Academy of Science
(NRC, 1983). Each of these four steps is described briefly according to CAC definitions.

Hazard Identification. Hazard identification involves identification of biological,
chemical, and physical agents capable of causing adverse health effects and that may be
present in a particular food or group of foods.

Hazard Characterization. Hazard characterization is the qualitative and/or quantitative
evaluation of the nature of the adverse health effects associated with biological, chemical,
and physical agents that may be present in food. Hazard characterization may or may not
include dose-response assessment.

7
Exposure Assessment. Exposure assessment is the qualitative and/or quantitative
evaluation of the likely intake of biological, chemical, and physical agents via food as
well as exposures via other sources, if relevant.

Risk Characterization. Risk characterization involves qualitative and/or quantitative
estimation, including attendant uncertainties, of the probability of occurrence and severity
of known or potential adverse health effects in a given population based on hazard
identification, hazard characterization, and exposure assessment.
1.4 Important Issues in Food Safety Modeling.
There are many important issues in modeling as applied to risk assessment that motivate
the need for sensitivity analysis and that illustrate the key challenges that sensitivity analysis
faces or that can be addressed by sensitivity analysis. Cullen and Frey (1999) discuss issues such
as purpose of models, model extrapolation, inappropriate application of models, complex
models, and model verification and validation.
1.4.1 Purpose of the Model.
The purpose of a model is to represent as accurately as necessary a system of interest. A
system is typically comprised of many components. All modeling involves decisions regarding
aggregation and exclusion. Aggregation refers to simplified representation of complex real
world systems. Exclusion refers to a decision to omit any portions that are judged not to be
important with respect to modeling objectives (Cullen and Frey, 1999).

8
The reason that risk models are developed is to serve as an aid in decision making.
Models are developed for different purposes, often with different decision making objectives in
mind. For example, three different objectives are addressed by screening analysis models,
research models, and assessment/decision making models, respectively.

Screening analysis is usually based on simple models that are not likely to underestimate
the risk. The purpose of this analysis is usually to help the decision maker with routine
regulatory decisions. Screening methods are intended to help identify exposure pathways that
are not important and to do so with a great deal of confidence. Screening methods cannot be
used to prove that a particular exposure pathway is important. Rather, they can identify exposure
pathways that should receive priority for further evaluation and analysis. Because screening
models are conservative, they are designed to provide some "false positives." This means that
they will often provide results indicating that a exposure pathway is of concern, but later analysis
with more refined assumptions or models may reveal less of a problem. A key advantage of
screening models is that they are often easier to use than more refined models. Screening models
typically have fewer input data requirements and involve less complex calculations.

Research models are intended to improve understanding of the function and structure of
real systems. They allow a researcher to explore possible and plausible functional relationships
and may involve many detailed mechanisms. Research models may be complicated and may not
be aimed at any particular risk management end point. As such, they may have large input data
requirements, be difficult to execute, and provide output not directly relevant to a specific risk
9
management objective. However, such models can be very helpful in improving fundamental
insights regarding risk processes and in learning lessons useful for further model development.

Refined assessment models are intended to serve as tools for decision making, such as for
rule-making or regulatory compliance purposes where screening analyses are inadequate. These
models are generally complex and more accurate than screening models. Refined models are
developed with at least some attention to risk management application. Compared to screening
models, refined models typically have greater data requirements and require more time and
experience to apply.

Sensitivity analysis is important for all three types of models. In the context of this
discussion, refined risk models are perhaps the most relevant. Sensitivity analysis can be used to
evaluate how robust risk estimates and management strategies are to model input assumptions
and can aid in identifying data collection and research needs.
1.4.2 Complex Models.
A model can be small yet complex. Model size should not be confused with model
complexity. Complex systems are often hierarchies, which can be described in terms of the
"span" of each level in the hierarchy and the number of levels (Reed and Afjeh, 2000; Simon,
1996).

Food Safety Risk Models (FSRMs) typically: (1) are nonlinear; (2) contain discrete
inputs; (3) contain thresholds; (4) contain multiple pathways; and (5) are modular. The nonlinear
and threshold features imply that interactions are important. For example, if temperature is low
10
enough, then there may be no microbial growth, but once temperature exceeds a threshold, then
the model may predict a nonlinear response that depends on several model inputs. The
modularity feature means that computations may take place in separate modules of the models,
and only a selected set of aggregated results may be passed from one module to another.
1.4.3 Model Verification and Validation.
Model verification is a process of making sure that the model is doing what it is intended
to do. Sensitivity analysis can be helpful in verification. If a model responds in an unacceptable
way to changes in one or more inputs, then troubleshooting efforts can be focused to identify the
source of the problem.

Model validation ideally involves comparison of model results to independent
observations from the system being modeled. Generally, in most risk assessments, complete
validation is not possible because of lack of sufficient observational data. Risk assessment
models usually predict rare events. For example, in radiation exposure assessments, model
validation may be nearly impossible due to extremely low levels of radionuclide concentrations
in the environment or because the time periods considered by the model are prohibitively long
(Hoffman and Miller, 1983). Cullen and Frey (1999) discuss partial validation of a model when
observational data are available for only a part of the modeling domain.

Sensitivity analysis can be used to help develop a "comfort level" with a particular
model. If the model response is reasonable from an intuitive or theoretical perspective, then the
model users may have some comfort with the qualitative behavior of the model even if the
quantitative precision or accuracy is unknown.
11
1.4.4 Model Extrapolation.
A model is applicable within the specific set of inputs and outputs associated with the
data used for calibration or validation of the model. Model extrapolation involves making
predictions for a situation beyond the range of calibration or validation of the model. For
example, Guassian-plume-based air dispersion models are applicable only for distances up to
about 20 km or 50 km from the emission source. The model should not be used to characterize
long-range transport over hundreds of kilometers.

Empirical models such as ones developed from regression analysis are generally based on
a specific data set. Regression analysis may be based on arbitrary functions to approximate
relationships between the data. However, there may not be any theoretical basis for the
functional relationship selected. In such a situation, if the inputs used are outside the original
data set from which the regression model was developed then the output may not valid. There
are two types of extrapolations. Explicit extrapolation involves making predictions for values of
inputs outside the range of values for which the model was calibrated or validated. Hidden
extrapolation involves specifying combinations of inputs for which validation has not been done,
even though each of the input values fall within the range of values that have been tested. If the
functional form of the relationship between the output and the inputs of the model is based on
sound theoretical assumptions, then the model may perform reasonably well when extrapolated.
Sensitivity analysis can be used to reveal how the model performs when the model is
extrapolated.
12
1.5 Sensitivity Analysis Methods.
Sensitivity analysis methods can be classified in a variety of ways. In this report, they are
classified as: (1) mathematical; (2) statistical; and (3) graphical. Other classifications focus on
the capability, rather than the methodology, of a specific technique (e.g., Saltelli et al., 2000).
Classification schemes aid in understanding the applicability of a specific method to a particular
model and analysis objective. Here, the focus is on sensitivity analysis techniques applied in
addition to the fundamental modeling technique. For example, an analyst may perform a
deterministic analysis, in which case a mathematical method, such as nominal range sensitivity
can be employed to evaluate sensitivity. Alternatively, an analyst may perform a probabilistic
analysis, using either frequentist or Bayesian frameworks, in which case statistical-based
sensitivity analysis methods can be used (e.g., Cullen and Frey, 1999; Box and Tiao, 1992;
Saltelli et al., 2000; Weiss, 1996).
1.5.1 Mathematical Methods for Sensitivity Analysis.
Mathematical methods assess sensitivity of a model output to the range of variation of an
input. These methods typically involve calculating the output for a few values of an input that
represent the possible range of the input (e.g., Salehi et al., 2000). These methods do not address
the variance in the output due to the variance in the inputs, but they can assess the impact of
range of variation in the input values on the output (Morgan and Henrion, 1990). In some cases,
mathematical methods can be helpful in screening the most important inputs (e.g., Brun et al.,
2001). These methods also can be used for verification and validation (e.g., Wotawa et al.,
1997) and to identify inputs that require further data acquisition or research (e.g., Ariens et al.,
13
2000). Mathematical methods evaluated here include nominal range sensitivity analysis, break-
even analysis, difference in log odds ratio, and automatic differentiation.
1.5.2 Statistical Methods for Sensitivity Analysis.
Statistical methods involve running simulations in which inputs are assigned probability
distributions and assessing the effect of variance in inputs on the output distribution (e.g.,
Andersson et al., 2000; Neter et al., 1996). Depending upon the method, one or more inputs are
varied at a time. Statistical methods allow one to identify the effect of interactions among
multiple inputs.

The range and relative likelihood of inputs can be propagated using a variety of
techniques such as Monte Carlo simulation, Latin hypercube sampling, and other methods.
Sensitivity of the model results to individual inputs or groups of inputs can be evaluated by
variety of techniques (Cullen and Frey, 1999). Greene and Ernhart (1993), Fontaine and
Jacomino (1997), and Andersson et al. (2000) give examples of the application of statistical
methods. The statistical methods evaluated here include regression analysis, analysis of
variance, response surface methods, Fourier amplitude sensitivity test, and mutual information
index.
1.5.3 Graphical Methods for Sensitivity Analysis.
Graphical methods give representation of sensitivity in the form of graphs, charts, or
surfaces. Generally, graphical methods are used to give visual indication of how an output is
affected by variation in inputs (e.g., Geldermann and Rentz, 2001). Graphical methods can be
used as a screening method before further analysis of a model or to represent complex
14
dependencies between inputs and outputs (e.g., McCamly and Rudel, 1995). Graphical methods
can be used to complement the results of mathematical and statistical methods for better
representation (e.g., Stiber et al., 1999; Critchfield and Willard, 1986).
1.6 Organization of the Report.
Section 2 identifies, describes, and evaluates ten selected sensitivity analysis methods.
Section 3 presents a comparison the methods. Conclusions are given in Section 4.

2.0 SENSITIVITY ANALYSIS METHODS.
This section identifies sensitivity analysis methods used across various disciplines.
Reference materials, such as journals, reports, books, and the World Wide Web were collected
from various fields including medical decision making, risk analysis, environmental science,
engineering, economics, ecology, statistics, health science, microbiology and food science.
Methods that were applied on models comparable to risk assessment models were identified for
evaluation. The methods and specific applications of each method are demonstrated. Strengths
and limitations of the methods are noted in brief.
2.1 Nominal Range Sensitivity.
Nominal range sensitivity method is also known as local sensitivity analysis or threshold
analysis (Cullen and Frey 1999; Critchfield and Willard, 1986). This method is applicable to
deterministic models. It is usually not used for probabilistic analysis. One use of nominal
sensitivity analysis is as a screening analysis to identify the most important inputs to propagate
15
through a model in a probabilistic framework (Cullen and Frey, 1999). Nominal range
sensitivity can be used to prioritize data collection needs as demonstrated by Salehi et al. (2000).
2.1.1 Description.
Nominal range sensitivity analysis evaluates the effect on model outputs exerted by
individually varying only one of the model inputs across its entire range of plausible values,
while holding all other inputs at their nominal or base-case values (Cullen and Frey, 1999). The
difference in the model output due to the change in the input variable is referred to as the
sensitivity or swing weight

of the model to that particular input variable (Morgan and Henrion,
1990). The sensitivity also can be represented as a positive or negative percentage change
compared to the nominal solution. The sensitivity analysis can be repeated for any number of
individual model inputs.

The results of nominal range sensitivity are most valid when applied to a linear model. In
such cases, it would be possible to rank order the relative importance of each input based upon
the magnitude of the calculated sensitivity measure as long as the ranges assigned to each
sensitive input are accurate. However, for a non-linear model, the sensitivity of the output to a
given input may depend on interactions with other inputs, which are not considered. Thus, the
results of nominal range sensitivity are potentially misleading for nonlinear models.
2.1.2 Application.
Two examples of the use of nominal range sensitivity analysis are given here. In the first,
nominal range sensitivity was applied as described here to an exposure assessment problem for
flounder in a contaminated harbor (Dakins et al., 1994). The selection of inputs evaluated in the
16
sensitivity analysis, and the range of values assigned to those inputs, was based upon literature
review. The absolute different in burden of PCB in flounder was the key measure of sensitivity.
An absolute difference of less than 0.5 g PCB per gram of body weight was considered to be
non-sensitive.

In the second example, a conditional nominal range sensitivity analysis was performed
(Sinicio et al., 1997). The objective of the analysis was to determine the allowable safe storage
time for grain wheat as a function of grain moisture content and temperature conditioned upon
key assumptions of a ventilation system. The intent was to recommend airflow rates and fan
control strategies to properly aerate the stored grains. The sensitivity analysis was conditioned
on two different assumptions regarding the airflow rate and two different assumptions regarding
the air temperature rise in the fan, leading to four different cases. For each case, three specific
inputs were each varied over a nominal range, and the impact on the key model output, allowable
storage time elapsed, was evaluated. Specific heat of the wheat was observed to be the most
sensitive input for each of the four cases.
2.1.3 Advantages.
Nominal range sensitivity analysis is a relatively simple method that is easily applied. It
works well with linear models and when the analyst has a good idea of plausible ranges that can
be assigned to each selected input. The results of this approach can be used to rank order key
inputs only if there are no significant interactions among the inputs, and if ranges are properly
specified for each input.
17
2.1.4 Disadvantages.
Nominal sensitivity analysis addresses only a potentially small portion of the possible
space of input values, because interactions among inputs are difficult to capture (Cullen and Frey
1999). Conditional sensitivity analysis may be used to account for correlation between inputs or
nonlinear interactions in model response, but it has limitations because of the combinational
explosion of possible cases. Potentially important combined effects on the decision (or output)
due to small changes in a few or all inputs together are not shown by nominal sensitivity analysis
For other than linear models, it is not clear that nominal range sensitivity can provide a reliable
rank ordering of key inputs.
2.2 Difference in Log-Odds Ratio ( LOR).
The difference in the log odds ratio (LOR) method is a specific application of nominal
range sensitivity methodology. The LOR is used when the output is a probability. For
example, Song et al. (2000) used LOR for identifying the most important inputs so that
additional data collection and research can be prioritized.
2.2.1 Description.
The odds or odds ratio of an event is a ratio of the probability that the event occurs to the
probability that the event does not occur (Walpole and Myers, 1993). If an event has a
probability of occurrence as P, then the odds ratio is P/(1-P). The log of the odds ratio or logit is
just another convenient way of rescaling probabilities (Menard, 1995). Log odds are considered
by some to be the preferred transformation of probability because they are putatively easier to
understand (Christensen, 1990).
18

The LOR method is used to examine the change in the output as:
1
]
1
1
]
1

) | Pr(
) | Pr(
log
) | Pr(
) | Pr(
log
changes without event No
changes without event
input in changes with event No
input in changes with event
LOR .
If LOR is positive, changes in one or more inputs enhance the probability of the specified event.
If LOR is negative, then the changes in the inputs cause a reduction in the probability of the
event occurring or increase the probability of the event not occurring. The greater the magnitude
of LOR, the greater is the influence of the input (Stiber et al., 1999).
2.2.2 Application
The LOR method was used by Stiber et al. (1999) to identify key inputs to a model of
groundwater decontamination via reductive dechlorination. The model inputs referred to as
evidences, included site parameters such as temperature, pH, and whether various specific
chemicals were found to be present, such as oxygen, hydrogen, chloride, dichlorethene (DCE),
methane, and others. The key model output is the probability that anaerobic degradation of
trichloroethene (TCE) via reductive dechlorination is occurring. An expert judgment-based
approach was used to estimate the output probability. A total of 14 possible pieces of evidence
were considered in the analysis. The results of the LOR analysis of Stiber et al. (1999) are
given in Figure 1. For example, if there is no evidence that oxygen is present then there is
increased probability of anaerobic processes occurring. If chloride or DCE is present, then there
is positive evidence of the degradation of TCE. From the sensitivity analysis, it appears that
evidence regarding DCE is the most important determinant of anaerobic degradation of TCE,
19
because changes in evidence for this particular input lead to the largest differences in the log-
odds ratio.
2.2.3 Advantages.
The LOR method is a useful measure of sensitivity when the model output is a
probability (Menard, 1995).
The LOR method can only be used when the output is in terms of probability (Menard,
1995). It suffers from other drawbacks similar to nominal range sensitivity analysis. For
example, similar to nominal range sensitivity analysis, LOR cannot account for nonlinear
interactions between or among inputs. Similar to nominal range sensitivity analysis, the
significance of differences among the sensitivities can be difficult to determine for nonlinear
models and correlated inputs, making it potentially difficult to rank order key inputs.
2.3 Break-Even Analysis.
Break-even analysis is more of a concept than a specific method. Broadly speaking, the
purpose of break-even analysis is to evaluate the robustness of a decision to changes in inputs
(von Winterfeldt and Edwards, 1986).
2.3.1 Description.
Break-even analysis involves finding values of inputs that provide a model output for
which a decision maker would be indifferent among the two or more risk management options.
The combinations of values of inputs for which a decision maker is indifferent to the decision
20
options are known as switchover or break-even values. Then, in order to assess the robustness of
a choice between the options, one can evaluate whether the possible range of values of the model
inputs corresponds with only one of the two choices (Morgan and Henrion, 1990). Indifference
of a decision maker to the two choices is often represented by a break-even line or indifference
curve such as an iso-risk curve. Ambiguity regarding selecting a particular choice exists if the
uncertainty range associated with an output may correspond to either of the two or more possible
choices. Different options that result in equivalent levels of risk reduction also can be identified
so that a decision maker can evaluate these options. If there are more than two decision options,
the analysis can get complex (von Winterfeldt and Edwards 1986).
2.3.2 Application.
Break-even analysis is often used in economics for purposes such as budget planning
(Dillon, 1993). Break-even analysis has also found applications in several other fields such as
health care (Boles and Fleming, 1996). Kottas and Lau (1978) describe the concept of stochastic
break-even analysis. Starr and Tapiero (1975) explain the use of break-even analysis with
consideration of risk.

As an example of a breakeven analysis, consider a choice between two medical treatment
options: medication versus an operation. The patients valuation of the possible outcome of
each option is represented with a utility function, and the probability of success of the operation
is not known. Figure 2 illustrates the combination of values for the utility of the medication and
the probability of success of the operation that lead to a preference for medication, as in the
upper left portion of the graph, or for the operation, as in the lower right portion of the graph.
Along the indifference line, there is equal preference for both options. If the patient has a high
21
utility for the outcome of the medication, and believes that the probability of success of the
operation is low, then medication is clearly preferred. If the range of uncertainty in the
probability of success is large, and if the range of uncertainty encloses the indifference line, then
it will be ambiguous as to what treatment option is preferred. In the former case, a decision can
be made that is robust to uncertainty. In the latter case, more information is needed to reduce
uncertainty in one or more of the inputs in order to make a decision with a high degree of
confidence.
2.3.3 Advantages.
The switchover or break-even point guides further modeling and elicitation. If the range
of uncertainty regarding an input encloses the break-even point, then that input will be important
in making a decision; that is, there will be uncertainty regarding which decision to take. In such
a situation, further research can be directed so as to help the decision maker to narrow the range
of uncertainty and make a decision with more confidence. On the other hand, if the uncertainty
regarding an input does not enclose the break-even point then there will be high confidence
regarding the decision.
Break-even analysis is not a straightforward method to apply. Though it is a useful
concept, its application is increasingly complex as the number of sensitive inputs increases (von
Winterfeldt and Edwards, 1986). There also is not a clear ranking method to distinguish the
relative importance of the sensitive inputs.
22
2.4 Automatic Differentiation Technique.
The automatic aifferentiation (AD) technique is an automated procedure for calculating
local sensitivities for large models. In AD, a computer code automatically evaluates first-order
partial derivatives of outputs with respect to small changes in the inputs. The values of partial
derivatives are a measure of local sensitivity.
2.4.1 Description.
Most existing sensitivity analysis methods based upon differentiation, such as numerical
differential methods, have one or more of the following limitations: inaccuracy in the results,
high cost in human effort and time, and difficulty in mathematical formulation and computer
program implementation (Hwang et al., 1997). To overcome these limitations, AD techniques
were developed. AD is a technique to perform local sensitivity analysis and not a new method in
itself. In AD the local sensitivity is calculated at one or more points in the parameter space of
the model. At each point, the partial derivatives of the model output with respect to a selected
number of inputs is evaluated.

AD is implemented by pre-compilers that analyze the code of the complex model and
then add instructions needed to compute first or higher order derivatives in an efficient manner to
save computational time and reduce complexity (Kedem, 1980; Rall, 1980; Carmichael et al.,
1997). The resulting expanded code is then complied with a standard compiler so that code can
evaluate function values used in the model, outputs, and derivatives. Automatic Differentiation
in Fortran (ADIFOR) is one of the softwares used for implementation of AD (Bischof et al.,
1992, 1994 and 1996).
23
2.4.2 Application.
Automatic differentiation finds application in models that involve complex numerical
differentiation calculations such as partial derivatives, integral equations, and mathematical
series (Hwang et al., 1997). It is used in fields such as air quality (Carmichael et al., 1997),
aerodynamics (Issac and Kapania, 1997), mechanical structures (Ozaki et al., 1995), and others.
Christianson (1999) used AD for verification of part of a model.

As an example, Carmichael et al. (1997) applied ADIFOR to calculate the sensitivity of
ambient air ozone concentration to air quality model inputs representing initial conditions,
reaction rate constants, and others. For this purpose, dimensionless sensitivity coefficients were
calculated based upon the numerical partial derivative of ozone concentration with respect to a
selected input, divided by the ratio of the values of the concentration and the input:
S
i,j
=
,
_
,
_
j
i
j
i
C
(2-1)
where,
S
i,j
= Normalized local sensitivity coefficients for i
th
chemical specie and j
th
input;
C
i
= The absolute change in the output concentration of i
th
chemical specie;
j
= The absolute change in the j
th
input;
j
= Values of model inputs; and
C
i
= Concentration of a chemical specie i at a given time.

24
The normalized local sensitivity coefficients were calculated for a 5-day simulation for
two different meteorological scenarios. The estimated ozone concentration was found to be most
sensitive to the initial ozone concentrations in both scenarios, with normalized sensitivity values
of 0.082 and 0.054 for the marine and land scenarios, respectively. In contrast, the results were
relatively insensitive to the initial NO concentration in both cases, as revealed by a normalized
sensitivity coefficient of approximately 0.001. Lumped sensitivities were also reported to give
the global effect of a perturbation in a given input on the whole system. Lumped sensitivity of a
chemical specie is the summation of normalized local sensitivities at different time stages for that
chemical specie.
2.4.3 Advantages.
AD techniques, such as ADIFOR can be applied without having detailed knowledge of
the algorithm implemented in the model. ADIFOR does everything automatically once it is
appended with the main code (Bischof et al., 1992). AD is superior to finite difference
approximations of the derivatives because numerical values of the computed derivatives are
more accurate and computational effort is significantly lower (Bischof et al., 1992). Hwang et
al. (1997) observed CPU time saving of 57% by using AD for sensitivity analysis as compared to
using a traditional method. If the model is constantly modified and improved, then ADIFOR
provides a convenient tool to easily accommodate such necessary model changes, which can be
very difficult to do in the case of other techniques (Carmichael et al., 1997).
The availability of AD technique may be limited to specific computer languages such as
FORTRAN in the case of ADIFOR, requiring the user to provide FORTRAN code for the
25
model. Because AD is a local technique, it suffers from the limitations of nominal range
sensitivity analysis. Furthermore, unlike nominal range sensitivity analysis, the possible range of
values of the inputs is not considered. The accuracy for sensitivity results is conditioned upon
the numerical method used in the AD software. Also, for nonlinear models, the significance of
differences in sensitivity between inputs is difficult to determine, making the rank ordering of
key inputs potentially difficult. This method cannot be used if partial derivatives cannot be
evaluated locally.
2.5 Regression Analysis.
Regression analysis can be employed as a probabilistic sensitivity analysis technique as
demonstrated by Iman et al. (1985). Regression analysis serves three major purposes: (1)
description of the relation between variables; (2) control of predictor variables for a given value
of a response variable; and (3) prediction of a response based on predictor variables (Neter et al.,
1996; Sen and Srivastava, 1990).
2.5.1 Description.
A relation between inputs and the output should be identified prior to regression analysis
based on techniques such as scatter plots or upon understanding of the functional form of the
model. Methods, such as stepwise regression, can be used to automatically exclude statistically
insignificant inputs. The regression model may not be useful when extrapolating beyond the
range of values used for each input when fitting the model (Devore and Peck, 1996).

Regression analysis is most properly performed on an independent random sample of
data. The effect of inputs on the output can be studied using regression coefficients, standard
26
errors of regression coefficients, and the level of significance of the regression coefficients
(Devore and Peck, 1996; Steel et al., 1997; Sen and Srivastava, 1990). Regression analysis
typically involves fitting a relationship between inputs and an output such as this linear one:

Y
i
= o +
1
X
1,i
+
2
X
2,i
+ +
m
X
m,i
+
i
(2-2)

where,
Y
i
= i
th
output data point for i
th
input data points;
X
j,i
= i
th
input data point for the j
th
input;
j
= regression coefficient for the j
th
input; and
= error for the i

th
data point.

Each term in the regression model can have a different basis function, which can be linear
or nonlinear. For a linear model, the regression coefficient
j
, can be interpreted as the change in
output Y
i
when the input X
j,i
for a given value of j increases by one unit and the values of all
other inputs remain fixed (Devore and Peck, 1996). Therefore, regression coefficients can be
used as a form of nominal range sensitivity. The deviation of the prediction of the regression
model from the actual can be measured using the coefficient of multiple determination, R
2
. R
2
is
a measure of the amount of variance in the dependent variable explained by the model (Draper
and Smith, 1981). A key assumption of least squares regression analysis is that the residuals are
normally distributed.

Because the regression coefficients are estimated from a random sample of data, the
coefficients themselves are random variables. If the coefficient is not significantly different than
27
zero, then there is not a statistically significant linear relationship between the input and the
output (Draper and Smith, 1981). Conversely, if the coefficient is statistically significant, then
there is stronger evidence of sensitivity. To determine statistical significance, the standard error
of the regression coefficient is estimated. If the ratio of the value of the regression coefficient
divided by its standard error is greater than a critical value, then the coefficient is deemed to be
statistically significant. The critical value is determined based upon the desired significance
level (usually 0.05) and the degrees of freedom of the regression model (Devore and Peck,
1996). The magnitude of statistically significant regression coefficients can be used to help
determine the ranking of the inputs according to their sensitivity if the inputs or the coefficients
are normalized (or standardized) to remove dimensional effects (Neter et al., 1996; Iman et al.,
1985). Neter et al. (1996) has suggested a transformation called correlation transformation
which standardize the data such that the standardized regression coefficients fall between -1 and
1. The advantage of this transformation is that the round-off errors can be minimized and all
regression coefficients have the same unit and hence, regression coefficients can be compared on
equal basis

In the case of forward stepwise regression analysis, the incremental change in the R
2

values is an indication of the significance of sensitivity of the output to each newly introduced
input. However, in the case of dependent or correlated inputs, the problem of multicollinearity
can affect the robustness of the results of regression analysis (Neter et al., 1996). A few types of
nonlinear models can be transformed into linear models in the case of logistic regression
(Hosmer and Lemeshow, 1989). Generalized linear models (GLM) provide flexibility to use
28
correlated input data and non-normal error distributions. Logistic regression and Poisson
regression are examples of GLM (McCullagh and Nelder, 1989; Searle, 1987).
2.5.2 Applications.
Regression analysis as a sensitivity analysis method is applied in various fields such as
veterinary science (Taylor et al., 1992), social sciences (Barrett et al., 1986), food sciences (Cliff
et al., 1995), and food safety (Tienungoon et al., 2000). McCarthy et al. (1995) used logistic
regression for sensitivity analysis of a stochastic population model using logistic regression.

As one example, Helton et al. (1995) used stepwise regression analysis for sensitivity
analysis of a model for contamination of food pathways associated with a severe accident at a
nuclear power station. A probabilistic analysis was done and forward stepwise regression
analysis was used to help identify which of the input distributions contributed most to
uncertainty in the mean population dose. The importance of inputs was indicated by the order in
which inputs entered the regression model, the changes in R
2
values with the entry of successive
inputs in the regression model, and the standardized regression coefficients in the final regression
model. The stepwise regression analysis for an example output, the mean population dose, is
shown in Table I. The mean population dose is dominated by the variables PSMCS134,
TFMCS, and PSMI131. Collectively these three variables account for 86 percent of the observed
variation. The remaining seven variables contribute only an additional eight percent increase in
the explanatory power of the model, and thus are relatively insensitive. Regression coefficients
for all these variables were statistically significant and variables with insignificant regression
coefficients were excluded from the forward stepwise regression.
29
2.5.3 Advantages.
Regression techniques such as the ones discussed here allow evaluation of sensitivity of
individual model inputs, taking into account the simultaneous impact of other model inputs on
the result (Cullen and Frey, 1999). Other regression techniques, such as those based upon the
use of partial correlation coefficients, can evaluate the unique contribution of a model input with
respect to variation in a selected model output (Brikes and Dodge, 1993). Moreover, a rank
regression approach may also be used. In rank regression, the ranks of each input and the output
are used instead of the sample values. Rank regression can capture any monotonic relationship
between an input and the output, even if the relationship is nonlinear. Sample and rank
regression methods are discussed elsewhere, such as by Neter et al. (1996), Iman et al. (1985),
Brikes and Dodge (1993), and Kendall and Gibbons (1990).
The key potential drawbacks of regression analysis include: possible lack of robustness
if key assumptions of regression are not met, the need to assume a functional form for the
relationship between an output and selected inputs, and potential ambiguities in interpretation.

Regression analysis works best if each input is statistically independent of every other
input (Devore and Peck, 1996). Furthermore, the residuals of a least squares regression analysis
must be normally distributed and independent. If these conditions are violated, the results of the
analysis may not have a strict quantitative interpretation, but instead should be treated as
providing conceptual or qualitative insights regarding possible relationships.

30
The results of sample regression analysis can be critically dependent upon the selection
of a functional form for the regression model (Neter et al., 1996). Thus, any results obtained are
conditioned upon the actual model used.

Regression analysis can yield results that may be statistically insignificant or counter
intuitive (Neter et al., 1996). The lack of a clear finding may be because the range of variation
of that input was not wide enough to generate a significant response in the output. Thus,
regression results can be sensitive to the range of variation in the data used to fit the model and
may not always clearly reveal a relationship that actually exists.
2.6 Analysis of Variance.
Analysis of variance (ANOVA) is a model independent probabilistic sensitivity analysis
method used for determining whether there is a statistical association between an output and one
or more inputs (Krishnaiah, 1981). ANOVA differs from regression analysis in that no
assumption is needed regarding the functional form of relationships between inputs and the
outputs. Furthermore, categorical inputs and groups of inputs can be addressed.
2.6.1 Description.
Inputs are referred to as "factors" and values of factors are referred to as factor levels in
ANOVA. An output is referred to as a "response variable." Single-factor ANOVA is used to
study the effect of one factor on the response variable. Multifactor ANOVA deals with two or
more factors and it is used to determine the effect of interactions between factors. A qualitative
factor is one where the levels differ by some qualitative attribute, such as a type of pathogen or
geographic regions. Quantitative factor levels are continuous quantities, such as temperature or
31
amount of pathogens. Neter et al. (1996) and Winter et al. (1991) give detail applications and
terminology for ANOVA.

ANOVA is a nonparametric method that is used to determine if values of the output vary
in a statistically significant manner associated with variation in values for one or more inputs. If
the output does not have a significant association with variation in the inputs, then the variation
in the output is random. The exact nature of the relationship between the inputs and the output is
not determined by ANOVA. Although the F-test is generally used to evaluated the significance
of the response of the output to variation in the inputs, additional tests such as the Tukey test and
Scheff test can also be used, such as to evaluate the effect of different input value ranges
(Montgomery, 1997; Hochberb and Tamhane, 1987).

In ANOVA, it is assumed that the output is normally distributed. Diagnostic checks are
important to determine whether the assumptions of ANOVA are violated. If any key
assumptions are violated then there can be corrective measures to address the problem. The F
test is generally robust to deviations from these assumptions but substantial departures from
normality or large differences in the variances of the output can influence statistical test results
(Lindman, 1974). In the case of correlated inputs, the results of the F test may not be robust.
However, approaches such as principal component analysis to group correlated factors can be
used to address this problem (Kim and Mueller, 1978).
2.6.2 Application.
ANOVA finds broad application across various fields such as health risk assessment
(e.g., Pet-Armacost et al., 1999; Gerken et al., 2000), material testing (e.g., Golinkin et al.,
32
1997), food quality (Carlucci et al., 1999, Ashraf et al., 1999), microbiology (e.g., McElroy et
al., 2000), and microbial risk assessment (e.g., Marks et al., 1998).

As an example of multifactor ANOVA, Carlucci et al. (1999) applied ANOVA to
evaluate the effects of factors, such as the age at slaughter, storage temperature, and the storage
time on response variables such as odor, flavor, and texture of the lamb meat. Each response
variable was evaluated using a 0100 scale for different combinations of the inputs. The
importance of input values was evaluated in terms of the relative magnitude of F values. All of
the response variables were significantly affected by the age at slaughter. In addition, the
response variable for texture of the meat was also influenced by an interaction term, indicating
that the effect of storage time depends on the age at slaughter.
2.6.3 Advantages.
No assumption is needed regarding the type of underlying model and both continuous
and discrete inputs can be analyzed using ANOVA (Montgomery, 1997). The results of
ANOVA can be robust to departures from key assumptions, and additional techniques can be
employed to deal with issues such as multicollinearity.
ANOVA can become computationally intensive if there are a large number of inputs. If
this becomes a problem, a suggestion by Winter et al. (1991) is to try to reduce the number of
inputs analyzed by using some less computationally intensive method, such as nominal range
sensitivity analysis, to screen out insensitive inputs. If there is a significant departure of the
response variable from the assumption of normality, then the results may not be robust
33
(Lindman, 1974). Errors in the response variables due to measurement errors in the inputs can
result in biased estimates of the effects of factors. If the inputs are correlated, then the effect of
each individual input on the response variable can be difficult to assess (Neter et al., 1996),
unless methods such as principal component analysis are used.
2.7 Response Surface Method (RSM).
The Response Surface Method (RSM) can be used to represent the relation between a
response variable (output) and one or more explanatory inputs (Myers and Montgomery, 1995;
Neter et al., 1996; Khuri and Cornell, 1987). The RSM can be used in a probabilistic analysis
(Chun et al., 1996; Frey and Bharvirkar, 1998). The RSM can identify curvatures in the
response surface by accounting for higher order effects. The RSM is generally complex and
therefore, used in later stages of an investigation when a limited number of factors are under
investigation (Neter et al., 1996).
2.7.1 Description.
A Response Surface (RS) can be linear or nonlinear and is typically classified as first-
order or second-order (Myers and Montgomery, 1995). The second-order structure is used when
there are interactions terms between inputs. The amount of time and effort required to develop a
response surface is typically a function of the number of inputs included and the type of response
surface structure required. It is often advantageous to limit the number of inputs that are
included in the response surface to those that are identified as most important using a screening
sensitivity analysis method, such as nominal range sensitivity analysis. The model must be
exercised for various desired combinations of the selected input values in order to generate a data
set that can be used to fit or calibrate a response surface. Monte Carlo simulation methods are
34
typically used to generate multiple values of each model input and to calculate corresponding
values of the model output.

A typical approach to response surface development is to use a least-squares regression
method to fit a standardized first or second order equation to the data obtained from the original
model. The key assumptions of least-squares regression (e.g., normality of the residuals) should
be reasonably satisfied; otherwise, other techniques such as rank-based or nonparametric
approaches, should be considered (Khuri and Cornell, 1987; Vidmar and McKean, 1996). The
precision and accuracy of the response surface can be evaluated by comparing the predictions of
the response surface to that of the original model for the same values of the model inputs. If the
precision and accuracy is not satisfactory, an improved fit might be obtained by iterating on
values of parameters for the response surface (Gardiner and Gettinby, 1998).

Once a response surface has been developed, the sensitivity of the model output to one or
more of the selected inputs can be determined by: (1) inspection of the functional form of the
response surface; (2) statistical analysis if regression analysis was used to develop the response
surface; or (3) application of other sensitivity analysis methods to the response surface. The
response surface can be thought of as a "model of a model" with the advantage of being simpler
and faster to execute than the original model. Therefore, computationally intensive sensitivity
analysis methods, such as Mutual Information Index or others, may be more readily applicable to
the response surface than to the original model.
35
2.7.2 Application.
The RSM is often employed for optimization and product quality studies (Gardiner and
Gettinby, 1998). Moskowitz (1997) demonstrated the use of RSM to optimize properties of
cereals to maximize consumer acceptance. Hopperstad et al. (1999) applied RSM for reliability
analysis of an aluminum extrusion process. The RSM also can be used for sensitivity analysis of
stochastic systems (Williams et al., 1999).

For example, Williams et al. (1999) describe application of RSM for optimization and
sensitivity analysis of manufacturing process of a type of integrated circuit. The complex
computations required for simulation of the manufacturing process previously limited the
strategies for process optimization. Therefore, a RS was incorporated into the simulator software
to enable statistical optimization and to perform sensitivity analysis. The RS included five of the
twenty original model inputs and was second order in structure.

Frey and Bharvirkar (1998) used the RSM to simplify a simulation model of a coal
gasification system. The RS models were developed from the data generated by probabilistic
simulation of the gasification process. Using regression techniques, a RS model was fitted to the
data. A series of equations were developed to enable prediction of 60 model outputs based upon
combinations of 12 inputs. These RS models were implemented with a FORTRAN-based
capital, operating, and levelized cost model to enable simultaneous prediction of the
performance, emissions, and cost of the complex technology of gasification using desktop
computers.
36
2.7.3 Advantages.
A key advantage of the RSM approach is that a potentially computationally-intensive
model can be reduced to a simplified form that enables much faster model run times. Therefore,
it will be easier to apply iterative numerical procedures to the RS, such as optimization or Monte
Carlo simulation, compared to the original model. Furthermore, the functional form of the RS
model and the values of its coefficients may provide a useful indication of key sensitivities.
Nominal range sensitivity or other methods can be applied to the RS model to elucidate
sensitivities, with faster run times.
In order to develop a RS, calculations with the original model are needed, which can be
resource intensive for some models. Because the RS is calibrated to data generated from the
original model, the valid domain of applicability of the RS model will be limited to the range of
values used to generate the calibration data set. Most RS studies are based on fewer inputs than
the original model. Therefore, the effect of all original inputs on the sensitivities cannot be
evaluated in RSM.
2.8 Fourier Amplitude Sensitivity Test.
The Fourier Amplitude Sensitivity Test (FAST) method is a procedure that can be used
for both uncertainty and sensitivity analysis (Cukier et al., 1973, 1975, and 1978). The FAST
method is used to estimate the expected value and variance of the output, and the contribution of
individual inputs to the variance of the output (Cukier et al., 1973). The FAST method is
independent of any assumptions about the model structure, and works for monotonic and non-
37
monotonic models (Saltelli et al., 2000). The effect of only one input (local sensitivity) or the
effect of all inputs varying together can be assessed by FAST.
2.8.1 Description.
The main feature of the FAST method is a pattern search method that selects points in the
input parameter space, and which is reported to be faster than the Monte Carlo method (McRae
et al., 1982). The classical FAST method is not efficient to use for high-order interaction terms
(Saltelli and Bolado, 1998). However, the extended FAST method developed by Saltelli et al.
(1999) can address higher order interactions between the inputs. Sobol's sensitivity method is
similar to the FAST method and can account for interacting terms, but it is less efficient than
extended FAST (Sobol, 1993).

A transformation function is used to convert values of each model input to values along a
search curve. As part of the transformation, a frequency must be specified for each input. By
using Fourier coefficients, the variance of the output is evaluated (Cuckier et al., 1973). The
contribution of input x
i
to the total variance is calculated based on the Fourier coefficients,
fundamental frequency
i
, and higher harmonics of the frequency as explained by Cuckier et al.
(1975). The ratio of the contribution of each input to the output variance and the total variance
of the output is referred to as the first order sensitivity index and can be used to rank the inputs
(Saltelli et al., 2000). The first order indices correspond to the contribution of individual inputs
and not to the contribution of interactions among inputs. To account for the residual variance in
the output due to higher order or interaction terms that is not explained by first order indices, the
extended FAST method is used (Saltelli et al., 1999).
38
The model needs to be evaluated at sufficient number of points in the input parameter
space such that numerical integration can be used to determine the Fourier coefficients (Saltelli
et al., 2000). The minimum sample size required to implement FAST is approximately eight to
ten times the maximum frequency used. In the case of discrete inputs, if a sufficiently large
sample size is not available, then the output can have frequent discontinuities. In such a case, the
Fourier coefficients may not be estimated properly and hence, the reliability of the results can be
adversely affected. Sobol's method is capable of handling discrete inputs (Saltelli et al., 2000).

McRae et al. (1982) describe mathematical basis and computer implementation of the
FAST method. Cukier et al. (1978) and Saltelli et al., (2000) give details of producing optimal
frequency sets. Different search curves and their transformation functions used in FAST are
given by McRae et al. (1982) and Cukier et al. (1975).
2.8.2 Application.
FAST has been applied in fields such as performance assessment of waste disposal
systems (e.g., Lu and Mohanty, 2001; Helton, 1993), atmospheric modeling (e.g., Rodriguez-
Camino and Avissar, 1998; Collins and Avissar, 1994; Liu and Avissar, 1996), and ground water
modeling (Fontaine et al., 1992).

As an example, Lu and Mohanty (2001) used the FAST method for sensitivity analysis of
a model developed for performance assessment of a proposed nuclear waste repository. The
model output is the amount of radiation for long time periods. Because the number of inputs of
the model is too large to be handled by the FAST method, less important input parameters were
first screened out. FAST was implemented using twenty inputs. For a 10,000 year time period
39
of interest, the top three most important inputs identified using FAST were thermal conductivity
of the rock material, the alluvium retardation coefficient for technetium, and the well pumping
rate for the farming receptor group located at 20 km. Conditional complementary cumulative
distribution functions of the model output (Mohanty and McCartin, 1998) were used to verify the
ranking of the influential inputs produced by the FAST method. The ranking of top three inputs
was found to be robust but the FAST method could not consistently rank other inputs of the set.
2.8.3 Advantages.
The FAST method is superior to local sensitivity analysis methods because it can
apportion the output variance to the variance in the inputs. It also can be used for local
sensitivity analysis with little modification (Fontaine et al., 1992). It is model independent and
works for monotonic and non-monotonic models (Saltelli et al., 2000). Furthermore, it can allow
arbitrarily large variations in input parameters. Therefore, the effect of extreme events can be
analyzed (e.g., Lu and Mohanty, 2001; Helton, 1993). The evaluation of sensitivity estimates
can be carried out independently for each factor using just a single set of runs (Saltelli et al.,
2000). The FAST method can be used to determine the difference in sensitivities in terms of the
differing amount of variance in the output explained by each input and, thus, can be used to rank
order key inputs.
The FAST method suffers from computational complexity for a large number of inputs
(Saltelli and Bolado, 1998). The classical FAST method is applicable to models with no
important or significant interactions among inputs (Saltelli and Bolado, 1998). However, the
extended FAST method developed by Saltelli et al., (1999) can account for high-order
40
interactions. The reliability of the FAST method can be poor for discrete inputs (Saltelli et al.,
2000).
2.9 Mutual Information Index.
The objective of the Mutual Information Index (MII) sensitivity analysis method is to
produce a measure of the information about the output that is provided by a particular input. The
sensitivity measure is calculated based upon conditional probabilistic analysis. The magnitude
of the measure can be compared for different inputs to determine which inputs provide useful
information about the output. MII is a computationally intensive method that takes into account
the joint effects of variation in all inputs with respect to the output. MII is typically used for
models with dichotomous outputs, although it can also be used for outputs that are continuous
(Critchfield and Willard, 1986).
2.9.1 Description.
The MII method typically involves three general steps: (1) generating an overall
confidence measure of the output value; (2) obtaining a conditional confidence measure for a
given value of an input; and (3) calculating sensitivity indices (Critchfield and Willard, 1986).
The overall confidence in the output is estimated from the CDF of the output. Confidence is the
probability for the outcome of interest. For example, if the dichotomous output is whether risk is
acceptable, the confidence is the probability that the risk is less than or equal to an acceptable
level. Conditional confidence is estimated by holding an input constant at some value and
varying all other inputs. The resulting CDF of the output indicates the confidence in the output
conditioned on a particular value of the input. The mutual information between two random
variables is the amount of information about a variable that is provided by the other variable
41
(Jelinek, 1970). The MII for each input is calculated based on the distribution of the input and
on both the overall and conditional confidence in the output. The average mutual information
index for an input is given by (Critchfield and Willard, 1986):

I
a
XY
=
x

y
P
X
P
Y|X
log
n
(P
Y | X
/ P
Y
) ( 2-3)
where,
P
Y|X
= conditional confidence;
P
Y
= overall confidence;
P
X
= probability distribution for the input; and
n = 2, to indicate binary output.

I
a
XY
is always positive. If I
a
XY
is large, then X provides a great deal of information about
Y. If X and Y are statistically independent, then I
XY
is zero. The amount of information about a
variable that is provided by the variable itself is measured in terms of the "average self
information" (I
YY
) of that variable, also known as the entropy of that variable (Jelinek, 1970).
The Average Self Information of the output Y is given as:
I
YY
=
Y
P
Y
log
n
(1 / P
Y
) ( 2-4)
where P
Y
is the overall confidence measure of Y. For the purpose of sensitivity analysis, a
normalized measure of the MII, the "natural sensitivity index," S
XY
is used.
S
XY
= (I
a
XY
/ I
YY
) X 100 % ( 2-5)

S
XY
reflects the percentage of the average mutual information contribution to the model
output Y that can be attributed to the input variable X. The calculation of MII may require
42
simplifying approximations regarding the use of a limited number of input values to represent the
variation in an input and estimation of probabilities of the input values. No analytical statistical
measure is available to determine the significance of the sensitivity indices. Because of the
assumptions and simplifications made in evaluating S
XY
, the robustness of rank ordering of key
inputs can be difficult to evaluate.
2.9.2 Application.
The MII method was devised by Critchfield and Willards (1986), who demonstrated its
application on a decision tree model. A dichotomous model was used to decide between two
options to treat the disease of deep vein thrombosis (DVT): anticoagulation and observation.
Each of these options had a relative importance to the decision maker, and they were valued in
terms of utility. The difference in the utilities (U) between the two options was used determine
the preferred option. The probability of a particular value of U for an option was interpreted as
the overall confidence level for that option. The overall confidence level for anticoagulation was
found to be 0.92 (92%).

In the conditional confidence analysis, an input of interest, such as the probability of
pulmonary embolism (PE) given DVT, P(PE|DVT) was held constant at one value and other
inputs were varied according to their respective distributions. The conditional confidence on U
for anticoagulation was determined. This process was repeated for all values of the input. The
results of the conditional confidence analyses were represented graphically. The mutual
information index and natural sensitivity index, S
XY
, were calculated. S
XY
for the input variable,
43
P(PE|DVT) was 25. The input variable, P(fetopathy), had a sensitivity index of 3.6. These
results indicated that the model output U was more sensitive to P(PE|DVT) than P(fetopathy).
2.9.3 Advantages.
MII includes the joint effects of all the inputs when evaluating sensitivities of an input.
The mutual information is a more direct measure of the probabilistic relatedness of two random
variables than other measures such as correlation coefficients (Jelinek, 1970). For example, the
correlation coefficient of two random variables examines the degree of linear relatedness of the
variables. Although two uncorrelated variables may not be independent, two variables with zero
mutual information are statistically independent. Therefore, the MII is a more informative
method. The results can be presented graphically facilitating their comprehension.
Calculation of the MII by Monte Carlo techniques suffers from computational
complexity, making practical application difficult (Merz et al., 1992). Critchfield and Willard
(1986a) have suggested an approach using symbolic algebra, which is reported to be less
computationally intensive. Because of the simplifying approximations that may be used in MII,
the robustness of ranking based on the sensitivity measure can be difficult to evaluate.
2.10 Scatter Plots.
A scatter plot is a graphical sensitivity analysis method. Scatter plots are used for visual
assessment of the influence of individual inputs on an output (Cook, 1994; Galvao et al., 2001).
A scatter plot is a method often used after a probabilistic simulation of the model. Scatter plots
44
are also often used as a first step in other analyses such as regression analysis and response
surface methods.
2.10.1 Description.
Each realization in a probabilistic simulation, such as a Monte Carlo simulation,
generates one pair of an input value and the corresponding output value. These simulated pairs
can be plotted as points on a scatter plot. Scatter plots depict the possible dependence between
an input and the output. Dependence may be linear or nonlinear (Cook, 1994). The range of
variation of the output may be constant regardless of the specific value of the input, or it may be
non-constant. Scatter plots may be based upon sample values, or based upon the ranks of the
values. General trends that might be observed in scatter plots are described by Kazmierski
(1995) and Bernstein et al. (1988).

Because scatter plots can help in visualizing and identifying potentially complex
dependencies between an input and an output, they can be used to guide the selection of
appropriate sensitivity analysis methods. For example, if the relationship is nonlinear, then a
nonlinear regression model, or a transformation of the data, may be required.

The number of data points displayed in a scatter plot needs to be selected such that there
is enough density to be able to observe any pattern, but not so many points that the variability
within the scatter is difficult to observe (Vose, 2000).

Kleijnen and Helton (1999) describe a sensitivity analysis method based on pattern
detection used to identify relationships between inputs and an output. For example, pattern
45
detection methods can identify linear and monotonic relationships between inputs and an output
as well as the effect of marginal distributions of an input and an output. Shortencarier and
Helton (1999) describe a pattern detection method and a related software tool.
2.10.2 Application.
Scatter plots have been used as an aid to sensitivity analysis in various fields such as
behavioral studies (e.g., Sobell, 1982; Rossier et al., 2001), environmental pollution (e.g.,
Hagent, 1992; Fujimoto, 1998), plant safety (e.g.,Helton et al., 2000), medical science (e.g.,
Bruno et al., 1994; Kogo and Ariel, 1997), veterinary science (e.g., Glaser, 1996), and others.

As an example, Moskowitz (1997) used consumer-based product evaluation to identify
optimal product formulations for a ready-to-eat cereal. Several physical attributes of the cereal,
such as appearance, color, flavor and quality which depend upon parameters (or inputs) such as
the amount of sweeteners and starch, die size, roasting time, and others, impact the customer
liking of the cereal. Thirty-one different product mixes were prepared by using different
combinations of eight such parameters. Different panels of consumers and experts tested these
mixes and used a 0 to 100 scale to rate their liking for the attributes of the products. Scatter plots
were used to represent the relationship between different combinations of the parameters (or
inputs) and the overall ratings of the cereal attributes (output).
2.10.3 Advantages.
Scatter plots are often recommended as a first step in sensitivity analysis of a statistical
sample of data, whether it is an empirical sample or the result of a probabilistic simulation. A
key advantage of scatter plots is that they allow for the identification of potentially complex
46
dependencies. An understanding of the nature of the dependencies between inputs and an output
can guide the selection of other appropriate sensitivity analysis methods.
A potential disadvantage of scatter plots is that they can be tedious to generate if one
must evaluate a large number of inputs and outputs unless commercial software is used to
automatically generate multiple scatter plots (SPLUS, 2000). Although not necessarily a
disadvantage, the interpretation of scatter plots can be qualitative and may rely on judgment.
Whether the sensitivities of two inputs differ significantly from each other cannot always be
judged from their scatter plots.

3.0 COMPARISON OF METHODS.
In this section, each of the ten methods are compared based upon four selected criteria:
(1) applicability of the method; (2) computational intensiveness; (3) ease and clarity in
representation of sensitivity; and (4) the purpose of the analysis. A summary of the comparison
for all ten methods is provided in Table II.
3.1 Applicability for Different Types of Models.
As indicated in Table II, some methods can be used only with deterministic models,
while others require the availability of probabilistic simulation results. Furthermore, some
methods are predicated upon an assumption of linearity, such as NRSA. Similarly, although AD
can be applied at multiple points in the input variable space, at each specific point it provides a
linear measure of sensitivity. Some methods are model dependent, such as regression analysis
47
and response surface methods, while the other methods do not require any assumption regarding
the functional form of the model. The LOR method and the AD method have specific
requirements regarding the form of the model output and regarding differentiability of the model,
respectively. The FAST method can be challenged by the presence of discrete inputs.

FSRMs are nonlinear, contain thresholds, and typically contain discrete inputs. At the
same time, FSRMs typically include probabilistic simulations of variability and/or uncertainty.
These characteristics suggest potential problems in application of NRSA and AD, which are
typically deterministic methods that assume linearity, and FAST, which may have difficulty with
discrete inputs in some cases. The large number of inputs to a FSRM may make break-even
analysis impractical; however, break-even could be applied to a FSRM after a small number of
most sensitive inputs have been identified using other methods. Methods such as ANOVA and
MII appear to be the most theoretically attractive methods, because they are model independent
and take into account the simultaneous interaction of all selected model inputs. However, MII in
particular is computationally intensive. Thus, it may be useful to develop a simplified response
surface based upon key inputs identified using less computationally intensive methods, and then
apply MII to the response surface.

The accuracy of a sensitivity analysis method will depend on whether all key
assumptions of the method hold true for a particular application. For example, if NRSA is used
with a model that is not only nonlinear, but also non-monotonic, it may be possible to obtain the
wrong insight. As an illustration, suppose that the value of a model output is the same for both
the low and high values of a given model input, but that the output has a maximum in the middle
48
of the range of values for the model input. NRSA will not reveal any relationship in this case.
Moreover, in the absence of any other analyses or insights by the analyst, NRSA will not by
itself reveal its own shortcoming in this case. In contrast, there are standard diagnostic tests that
can be used with methods such as regression analysis and ANOVA to determine if key
assumptions have been violated.
3.2 Computational Issues.
FSRMs can be large and computationally intensive. Therefore, there are incentives to
identify sensitivity analysis methods that do not introduce substantial additional computational
burdens. Some of the methods identified, including regression analysis, ANOVA, and scatter
plots, can be applied to the results of a probabilistic analysis without need for any additional
model runs. This can be useful only if all candidate sensitive variables have been assigned
distributions in the probabilistic analysis. Moreover, computational intensity is based not just on
exercise of the original model, but also upon any additional computations required in applying
the sensitivity analysis method. For example, the time to perform an ANOVA calculation will
be sensitive to the sample size, the number of inputs, and whether there are interactions among
the inputs. The time required to prepare scatter plots will depend on the number of pair-wise
combinations of inputs and outputs that are evaluated. Commercial software are readily
available in both of these examples, as well as for other methods such as regression analysis.

NRSA is potentially an easy method to apply if a limited number of inputs are to be
evaluated. In a FSRM, there may be some inputs that are not treated probabilistically which
could potentially be important in determining the model output. NRSA is in some ways
49
preferred over AD methods, because both the sensitivity and possible range of values are taken
into account in the former, while only local sensitivity is considered in the latter.

Aside from computational intensity, another consideration is computational feasibility.
AD methods can only be applied if the model is locally differentiable. FAST can encounter
difficulties with discrete model inputs. Thus, it is possible that either of these two methods may
be inapplicable to a particular model, unless the model or method can be revised to correct for
these issues.
3.3 Ease and Clarity in Representation of Sensitivity.
Each sensitivity analysis method has its own metric for sensitivity. There are standard
numerical ways of representing results for NRSA, LOR, AD, regression analysis, ANOVA,
FAST, and MII. For regression and ANOVA based methods, there are statistical measures of
the significance of the sensitivity indicators (e.g., regression model coefficients, F-ratios) that
can be used to help distinguish which inputs are insensitive from those that have at least some
influence on the output. Regression-based methods also provide information that might help
distinguish among the statistically significant inputs. For example, in standardized step-wise
regression, the magnitude of the regression coefficient and the incremental improvement in the
R
2
value can be compared for different inputs to gain insight into which ones are substantially
more sensitive than others. In some cases, a model output may have approximately equal
sensitivity to two or more inputs. Thus, small differences in the sensitivity indicators may be
insignificant from a practical, if not statistical, perspective. As a result, there may be ambiguity
50
regarding the rank order of inputs from most sensitivite to least sensitive with respect to a given
output.

For the other methods, such as break-even analysis, RSM, and scatter plots, the
representation of sensitivity may be less prescribed and more subject to individualized
presentation and interpretation. In the case of RSM, the interpretation of sensitivity may be
based upon examination of the functional form of the response surface, and the values of its
coefficients. If a statistical method, such as regression, was used to develop the response
surface, then there may be statistical measures of the significance of sensitivities between the
output and one or more inputs (including interaction terms). The interpretation of break-even
analysis results typically distills to whether or not a decision is sensitive to the range of
uncertainty in one or more inputs. Additional analysis might clarify how sensitive the decision
may be. For example, one could estimate the probability that a particular option is not the best
one depending on the distribution of uncertainty assigned to one of the key decision inputs. In
the case of scatter plots, the interpretation will often be case-specific. In some situations, clear
insights may emerge that enable identification of the most important inputs.

Sensitivity analysis methods can be distinguished based upon whether they are local or
global methods, and regarding whether they focus on one input at a time or take into account
simultaneous interactions among the inputs. A method that is global and that takes into account
interactions among multiple inputs may provide more robust insights than methods without these
features. NRSA, LOR, and AD are local methods and they do not account for simultaneous
interactions among the inputs. Break-even analysis may account for interactions among a small
51
number of inputs, conditioned on either point-estimates or distributions for all remaining inputs.
Thus, break-even analysis is a hybrid of both local and global characteristics. Regression,
ANOVA, FAST, MII, RSM, and scatter plots can be applied over a potentially large domain of
the input parameter space and can be applied taking into account simultaneous variation of
multiple inputs.
3.4 Purpose of the Analysis.
Sensitivity analysis may be performed for a variety of reasons, including identification of
key inputs and/or exploration of the model response to specific inputs. The latter can be useful
to either verification and/or validation of the model.

All of the methods reviewed here have the potential to reveal key sensitivities as long as
key assumptions for a given method are not grossly violated. Some methods provide more
information regarding the nature of the sensitivity than others. For example, because ANOVA is
a model-free and nonparametric approach, it only provides indication of a statistical association
between an output and an input. MII provides a measure of the information that an input imparts
to the output. FAST provides insight regarding the contribution of an input to the variance in the
output. None of these methods provide insight into the specific nature of the relationship. Thus,
these methods may be helpful in identifying key inputs, and in identifying qualitative features of
the relationship between an input and an output, but they are less useful in terms of quantitative
verification and/or validation of a model.

52
In contrast, a technique such as regression analysis or RSM can be used to help find a key
functional relationship that enables quantitative estimation of how the output changes with
respect to variation in the input. Methods such as NRSA, LOR, and AD provide insight
regarding the local behavior of the model, and thus have the potential to be useful for both
verification and validation. Scatter plots may in some sense be the most revealing of possible
problems in model performance, since no prior assumption is made regarding how the model is
responding to a given input.

Of course, with simple modifications, some methods can have increased utility as a
means for verification or validation of model performance. For example, a strict application of
NRSA may not take into account nonlinearities in model response. However, if three or more
values of a given model input are used to predict the model output, and if all such values are
displayed graphically, any nonlinearity in the model response can be revealed and compared with
prior expectations or with validation data.

4.0 CONCLUSIONS.
There are a large number of sensitivity analysis methods used in a variety of disciplines.
Only ten such methods have been identified and evaluated here. The methods were selected
based upon a judgment that they are widely used and of potential relevance to FSRMs. Each
method was characterized individually, and the methods were compared on the basis of four
criteria.

53
There is no single method that is clearly superior to all others. Each method has its own
key assumptions and limitations. Each method has its own demands regarding the time and
effort to apply the method and interpret the results. Each method has strengths and limitations
regarding the type of insight that it can provide. FSRMs have important features that, taken
individually, favor one method over another. However, when taken in aggregate, there is not one
obvious best method. For example, a model-free and global method is recommended as the most
appropriate choice because FSRMs typically are nonlinear, contain thresholds, and include
discrete inputs. Thus, methods such as NRSA, LOR, AD, regression analysis, and RSM may
not be able to provide robust insights either because they assume linearity or because they
require specification of a functional form. FAST, in spite of its many attractive features, may be
inapplicable because of potential problems in dealing with discrete inputs. In contrast, methods
such as ANOVA, MII, and scatter plots are model independent and in principle can deal with the
key characteristics of FSRMs. ANOVA can be applied to the results of a probabilistic analysis
but does not by itself provide direct insight into the nature of the relationship between the output
and the most sensitive inputs. MII would require computationally intensive simulations which
may be impractical unless a good response surface can be used instead of the original model.
Scatter plots should be used to help identify relationships that may otherwise be difficult to
characterize with other methods.

An iterative approach to sensitivity analysis is likely to be needed for most FSRMs. In
the iterative approach, one or more methods that are not computationally intensive can be used to
make a preliminary identification not only of the most important sensitive inputs, but also of the
least important inputs. Subsequently, more refined and/or computationally intensive sensitivity
54
analysis methods can be applied only to the subset of inputs that appear to be most important or
that a particular method can easily address.

Because each sensitivity analysis methods is typically based upon a different assumption
regarding appropriate ways of measuring sensitivity, it is possible that two different methods
may lead to different rank ordering of key inputs. Thus, a general recommendation is to use two
or more methods, preferably with dissimilar foundations, to increase confidence that the
identification of key inputs is robust.

Other important issues that must be addressed with FSRMs include modularity and
aggregation. FSRMs typically consistent of separate modules that represent different food
processing steps. Within a given part of a model, there may be aggregation (e.g., from individual
animals to batches of ground beef). As part of future work, methods should be identified,
applied, and evaluated for sensitivity analysis across modules and taking into account
aggregation.

Although there are theoretical arguments in favor of some methods over others, these
methods should be compared to evaluate whether their results differ in practice. Thus, for future
work, a quantitative comparison of multiple sensitivity analysis methods applied to the same
specific refined FSRMs is recommended. Such a comparison can provide insight regarding
whether the methods, in spite of different theoretical foundations, perform similarly in practice.

55
5.0 ACKNOWLEDGEMENT.
This work was supported by a cooperative agreement between the U.S. Department of
Agriculture's Office of Risk Asseessment and Cost Benefit Analysis and North Carolina State
University. The authors benefited from suggestions provided by participants of the June 11-12,
2001 NCSU/USDA Workshop on Sensitivity Analysis and from peer reviewer comments.

56
6.0 REFERENCES.
Agro, K.E., C.A. Bradley, and N. Mittmann (1997), "Sensitivity Analysis in Health Economic
and Pharmacoeconomic Studies - an Appraisal of the Literature," Pharmacoeconomics,
11(1):75-88.
Andersson, F.O., M. Aberg, and S.P. Jacobsson (2000), "Algorithmic Approaches for Studies of
Variable Influence, Contribution and Selection in Neural Networks," Chemo-metrics and
Intelligent Laboratory Systems, 51(1):61-72.
Ariens, G.A., W. van Mechelen P.M. Bongers, L.M. Bouter, and G. van der Wal (2000),
"Physical Risk Factors for Neck Pain," Scandinavian Journal of Work Environment &
Health, 26(1):7-19.
Ashraf, H.R., M. White, and B. Klubek (1999), "Microbiological Survey of Tofu Sold in a Rural
Illinois County," Journal of Food Protection, 62(9):1050-1053.
Baker, S., D. Ponniah, and S. Smith (1999), " Survey of Risk Management in Major UK
Companies," Journal of Professional Issues in Engineering Education and Practice,
125(3):94-102.
Baniotopoulos, C.C. (1991), "A Contribution to the Sensitivity Analysis of the Sea-Bed Structure
Interaction Problem for Underwater Pipelines," Computers & Structures, 40(6):1421-
1427.
Barrett, G.V., R.A. Alexander, M.N. Anesgart, and D. Doverspike (1986), "Frequently
Encountered Problems in the Application of Regression-Analysis to the Investigation of
Sex-Discrimination in Salaries," Public Personnel Management, 15(2):143-157.
57
Beck, M.B., J.R. Ravetz, L.A. Mulkey, and T.O. Barnwell (1997), "On the Problem of Model
Validation for Predictive Exposure Assessments," Stochastic Hydrology and Hydraulics,
11(3):229-254.
Bernstein, I.H., C.P. Garbin, and G.K. Teng (1988). Applied Multivariate Analysis. Springer-
Verlag: New York.
Bischof, C.H., A. Carle, G. Corliss, A. Griewank, and P. Hovland (1992). ADIFOR Generating
Derivative Codes From FORTRAN programs. ANL-4501, Prepared by Mathematics
and Computer Science Division, Aargonne National Laboratory, Argonne, Illinois.
Bischof, C.H., A. Carle, P. Khademi, and A. Mauer (1994), The ADIFOR2.0 System for the
Automatic Differentiation of FORTRAN77 Programs. IEEE Computational Science and
Engineering, 3(3):18-32.
Bischof, C.H., A. Carle, P. Khademi, and A. Mauer (1996), "ADIFOR 2.0: Automatic
Differentiation of Fortran 77 Programs," IEEE Computational Science & Engineering,
3(1):18-32.
Boles, K.E., and S.T. Fleming (1996), "Break-even Under Capitation: Pure and Simple?,"
Health Care Management Review, 21(1):38-47.
Box, G.E.P., and G.C. Tiao (1992). Bayesian Inference in Statistical Analysis. John Wiley and
Sons, Inc.: New York.
Brikes, D., and Y. Dodge (1993). Alternative Methods of Regression. John Wiley and Sons:
New York.
Brun, R., P. Reichert, and H.R. Kunsch (2001), "Practical Identifiability Analysis of Large
Environmental Simulation Models," Water Resources Research, 37(4):1015-1030.
58
Bruno, A., G. Delpoeta, A. Venditti, R. Stasi, G. Adorno, G. Aronica, G. Suppo, A.M. Dirienzo,
R. Iazzoni, M. Tribalto, T. Caravita, M. Masi, and G. Papa (1994), "Diagnosis of Acute
Myeloid-a System Coulter VCS," Haematologica, 79(5):420-428.
Buchanan, R.L. (1995), "The Role of Microbiological Criteria and Risk Assessment in HACCP,"
Food Microbiology, 12(5):421-424.
Buchanan, R.L. (1996), "Indentifying and Controlling Emerging Food Borne Pathogens:
Research Need," Emerging Infectious Diseases, 3(4):517-521.
Buchanan, R.L., Smith J.L., and W. Long (2000), "Microbial Risk Assessment: Dose-Response
and Risk Characterization," International Journal of Food Microbiology, 58(3):159-172.
Carlucci, A., F Napolitano, A. Girolami, and E. Monteleone (1999), "Methodological approach
to evaluate the effects of age at slaughter and storage temperature and time on sensory
profile of lamb meat," Meat Science, 52(4):391-395.
Carmichael, G.R., A. Sandu, and F.A. Potra (1997), Sensitivity Analysis for Atmospheric
Chemistry Models Via Automatic Differentiation, Atmospheric Environment, 31(3):475-
489.
Cheng, T.C.E. (1991), "EPQ with Process Capability and Quality Assurance Considerations,"
Journal of the Operational Research Society, 42(8):713-720.
Christensen, Ronald (1990). Log-Linear Models. Springer-Verlag: New York.
Christianson, B. (1999), "Cheap Newton Steps for Optimal Control Problems: Automatic
Differentiation and Pantoja's Algorithm," Optimization Methods and Software,
10(5):729-743.
59
Chun, M.H., S.J. Han, and N.I. Tak (1996), "A Combined Procedure of RSM And LHS for
Uncertainty Analyses in Source Term Quantifications Using Maap3.0B Code," Annals of
Nuclear Energy, 23(16):1337-1349.
Cliff, M.A., M.C. Dever, J.W. Hall, and B. Girard (1995), "Development and Evaluation of
Multiple Regression Models for Prediction of Sweet Cherry Liking," Food Research
International, 28(6):583-589.
Collins, D.C., and R. Avissar (1994), "An Evaluation with the Fourier Amplitude Sensitivity
Test (FAST) of Which Land-Surface Parameters Are of Greatest Importance in
Atmospheric Modeling," Journal of Climate, 7(5):681-703.
Cook, R.D (1994). An Introduction to Regression Graphics. John Wiley and Sons: New York.
Critchfield, G.C., and K.E. Willard (1986), "Probabilistic Analysis of Decision Trees Using
Mote Carlo Simulation," Medical Decision Making, 6(1):85-92.
Critchfield, G.C., and K.E. Willard (1986a), "Probabilistic Analysis of Decision Trees Using
Symbolic Algebra," Medical Decision Making, 6(1):93-100.
Cukier, R.I., C.M. Fortuin, K.E. Shuler, A.G. Petschek, and J.H. Schailby (1973), Study of the
Sensitivity of the Coupled Reaction Systems to Uncertainties in Rate Coefficients: I.
Theory, Journal of Chemical Physics, 59(8):3873-3878.
Cukier, R.I., H.B. Levine, and K.E. Shuler (1978), Nonlinear Sensitivity Analysis of
Multiparameter Model Systems, Journal of Computational Physical, 26(1):1-42.
Cukier, R.I., J.H. Schailby, and K.E. Shuler (1975), Study of the Sensitivity of Coupled
Reaction Systems to Uncertainties in Rate Coefficients: III. Analysis of
Approximations, Journal of Chemical Physical, 63(3):1140-1149.
60
Cullen, A.C., and H.C. Frey (1999). Probabilistic Techniques in Exposure Assessment. Plenum
Press: New York.
Dakins, M.E., J.E. Toll, and M.J. Small (1994), "Risk-Based Environmental Remediation:
Decision Framework and Role of Uncertainty," Environmental Toxicology and
Chemistry, 13(12):1907-1915.
Devore, J.L., and R. Peck (1996). Statistics: The Exploration and Analysis of Data. 3
rd
Edition.
Brooks/Cole Publishing Company: London, England.
Dillon, C.R. (1993), "Advanced Break-Even Analysis of Agricultural Enterprise Budgets,"
Agricultural Economics, 9(2):127-143.
Draper, N.R., and H. Smith (1981). Applied Regression Analysis. Second Edition. John Wiley
and Sons: New York.
ERS (2001), "Valuing the Health Benefits of Food Safety: A Proceedings." Report No.: ERS MP
1570. Economic Research Service, USDA.
http://www.ers.usda.gov/publications/mp1570/, Viewed: May 10, 2001.
FAO (1999), "CODEX Alimentarius Commission," Report No. ALINORM 99/13A. Food and
Agriculture Organization (FAO) of the United Nations.
http://www.codexalimentarius.net/, Viewed: May 10, 2001.
Fontaine, D.D., P.L. Havens, G.E. Blau, and P.M. Tillotson (1992), The Role of Sensitivity
Analysis in Groundwater Risk Modeling for Pesticides, Weed Technology, 6(3):716-
724.
61
Fontaine, T.A., and V.M.F. Jacomino (1997), "Sensitivity Analysis of Simulated Contaminated
Sediment Transport," Journal of the American Water Resources Association, 33(2):313-
326.
Fraedrich, D., A. Goldberg (2000), "A Methodological Framework for the Validation of
Predictive Simulations," European Journal of Operational Research, 124(1):55-62.
Frey, H.C., and R. Bharvirkar (1998), "Desktop Modeling of the Performance, Emissions, and
Cost of Gasification Systems, Proceedings of the 91st Annual Meeting, Paper No. 98-
RAA.02P, Air and Waste Management Association, Pittsburgh, Pennsylvania.
Fujimoto, K. (1998), " Correlation Between Indoor Radon Concentration and Dose Rate in Air
from Terrestrial Gamma Radiation in Japan," Health Physics, 75(3):291-296.
Galvao, L.S., M.A. Pizarro, and J.C.N. Epiphanio (2001), "Variations in Reflectance of Tropical
Soils: Spectral-Chemical Composition Relationships from AVIRIS Data," Remote
Sensing of Environment, 75(2):245-255.
Gardiner, W.P., and G. Gettinby (1998). Experimental Design Techniques in Statistical Practice:
A Practical Software-Based Approach. Horwood Publishing Ltd.: Chichester, West
Sussex, England.
Geldermann, J., and O. Rentz (2001), "Integrated Technique Assessment with Imprecise
Information as a Support for the Identification of Best Available Techniques (BAT)," OR
Spektrum, 23(1):137-157.
Gerken, M., L. Kreienbrock, J. Wellmann, M. Kreuzer, and H.E. Wichmann (2000), "Models for
Retrospective Quantification of Indoor Radon Exposure in the Case-Control Studies,"
Health Physics, 78(3):268-278.
62
Glaser, S. (1996), "Application of Statistical Classification Methods to the Evaluation of In Vitro
Tests," Alternatives to Laboratory Animals, 24(1):49-62.
Golinkin, I.A., D.D. Ruff, E.P. Kvam, G.P. McCabe, and A.F. Grandt (1997), "Application of
Analysis of Variance (ANOVA) Statistical Methods to Breaking Load Corrosion Test,"
Journal of Testing and Evaluation, 25(6):565-570.
Greene, T., and C.B. Ernhart (1993), "Dentin Lead and Intelligence Prior to School Entry - A
Statistical Sensitivity Analysis," Journal of Clinical Epidemiology, 46(4):323-339.
Hagent, N.T. (1992), " Macroparasitic Epizootic Disease: A potential Mechanism for the
Termination of Sea-Urchin Outbreaks in Northern Norway," Marine Biology,
114(3):469-478.
Hathaway, C.N. (1995), "Harmonization Requirements Under HACCP-Based Control Systems,"
Food Control, 6(5):267-276.
Helton, J.C. (1993), "Uncertainty and Sensitivity Analysis Techniques for Use in Performance
Assessment for Radioactive-Waste Disposal," Reliability Engineering and System Safety,
42(2-3): 327-367.
Helton, J.C., and R.J. Breeding (1993), "Evaluation of Reactor Accident Safety Goals,"
Reliability Engineering & System Safety, 39(2):129-158.
Helton, J.C., J.D. Johnson, J.A. Rollstin, A.W. Shiver, and J.L. Sprung (1995), Uncertainty and
Sensitivity Analysis of Food Pathway Results With the MACCS Reactor Accident
Consequence Model, Reliability Engineering and System Safety, 49(2), 109-144.
Helton, J.C., J.E. Bean, K. Economy, J.W. Garner, R.J. MacKinnon, J. Miller, J.D. Schreiber, P.
Vaughn (2000), "Uncertainty and Sensitivity Analysis for Two-Phase Flow in the
63
Vicinity of the Repository in the 1996 Performance Assessment for the Waste Isolation
Pilot Plant: Disturbed Conditions," Reliability Engineering and System Safety, 69(1-
3):263-304.
Hochberg, Y., and A.C. Tamhane (1987). Multiple Comparison Procedures. John Wiley and
Sons: New York.
Hoffman, F.O., and C.W. Miller (1983). "Uncertainties in Environment Radiological Assessment
Models and Their Implications," Proceedings of the Nineteenth Annual Meeting of the
National Council on Radiation Protection and Measurements. National Academy of
Science, Washington, DC.
Hopperstad, O.S., B.J. Leira, S. Remseth, and E. Tromborg (1999), Reliability-based Analysis
of a Stretch-bending Process for Aluminum Extrusions, Computers and Structures,
71(1):63-75.
Hosmer, D.W., and S. Lemeshow (1989). Applied Logistic Regression. John Wiley: New York.
Hwang, D., D.W. Byun, and M.T. Odman (1997), An Automatic Differentiation Technique for
Sensitivity Analysis of Numerical Advection Schemes in Air Quality Models,
Atmospheric Environment, 31(6):879-888.
Iman, R.L., M.J. Shortencarier, and J.D. Jhonson (1985). "A FORTRAN 77 Program and Users
Guide for the Calculation of Partial Correlation and Standardized Regression
Coefficients." Report No. SAND85-0044, Sandia National Laboratories, Albuquerque,
NM.
64
Issac, J.C., and R.K. Kapania (1997), Aero-elastic Sensitivity Analysis of Wings Using
Automatic Differentiation, Journal of American Institute of Aeronautics and
Astronautics, 35(3):519-525.
Jelinek, F. (1970). Probabilistic Information Theory. McGraw-Hill Book Company: New York.
Jones, R.N. (2000), "Analyzing the Risk of Climate Change Using an Irrigation Demand Model,"
Climate Research, 14(2):89-100.
Kaplan, S., and B.J. Garrick (1981), "On the Quantitative Definition of Risk," Risk Analysis,
1(1):11-27.
Kazmierski, T.J. (1995). Statistical Problem Solving in Quality Engineering. McGraw-Hill, Inc:
New York.
Kedem, G. (1980), " Automatic Differentiation of Computer Programs," ACM Transactions on
Mathematical Software, 6(2):150-165.
Kendall, M.G., and J.D. Gibbons (1990). Rank Correlation Methods. Fifth Edition. Charles
Griffin: London.
Kewley, R.H., M.J. Embrechts, and C. Breneman (2000), "Data Strip Mining for the Virtual
Design of Pharmaceuticals with Neural Networks," IEEE Transactions on Neural
Networks, 11(3):668-679.
Khuri, A.J., and J.A. Cornell (1987). Response Surfaces. Marcel Dekker, Inc.: New York.
Kim, J.O., and C.W. Mueller (1978). Introduction to Factor Analysis: What It Is and How to Do
It. Sage Publications: Beverly Hills, CA.
65
Kleijnen, J.P.C. (1995), "Verification and Validation of Simulation-Models," European Journal
of Operational Research, 82(1):145-162.
Kleijnen, J.P.C., and J.C. Helton (1999). "Statistical Analysis of Scatter Plots to Identify
Important Factors in Large-Scale Simulations." SAND98-2202, Prepared by Sandia
national Laboratories, Albuquerque, New Maxico.
Kleijnen, J.P.C., and R.G. Sargent (2000), " A Methodology for Fitting and Validating
Metamodels in Simulation," European Journal Of Operational Research, 120(1):14-29.
Kogo, N., and M. Ariel (1997), "Membrane Properties and Monosynaptic Retinal Excitation of
Neurons in the Turtle Accessory Optic System," Journal of Neurophysiology, 78(2):614-
627.
Kottas, J.F., and H.S. Lau (1978), "Stochastic Break-Even Analysis," Journal of the
Operational Research Society, 29(3):251-257.
Krishnaiah, P.R (1981). Ananlysis of Variance. Elsevier: New York.
Lammerding, A.M. (1997), "An Overview of Microbial Food Safety Risk Assessment," Journal
of Food Protection, 60(11):1420-1425.
Limat, S., M.C. Woronoff-Lemsi, E. Deconinck, E. Racadot, M. Jacquet, P. Herve, and J.Y.
Cahn (2000), "Cost-Effectiveness of Cd34(+) Dose in Peripheral Blood Progenitor Cell
Transplantation for Non-Hodgkin's Lymphoma Patients: a Single Center Study," Bone
Marrow Transplantation, 25(9):997-1002.
Lindman, H.R (1974). Analysis of Variance in Complex Experimental Designs. W. H. Freeman
& Co.: San Francisco.
66
Liu, Y.Q., and R. Avissar (1996), "Sensitivity of Shallow Convective Precipitation Induced by
Land Surface Heterogeneities to Dynamical and Cloud Microphysical Parameters,"
Journal of Geophysical Research-Atmospheres, 101(D3):7477-7497.
Lu, Y.C., and S. Mohanty (2001), "Sensitivity Analysis of a Complex, Proposed Geologic Waste
Disposal System Using the Fourier Amplitude Sensitivity Test Method," Reliability
Engineering and System Safety, 72(3):275-291.
Manheim, L.M. (1998), "Health Services Research Clinical Trials: Issues in the Evaluation of
Economic Costs and Benefits," Controlled Clinical Trials, 19(2):149-158.
Marks, H.M., M.E. Coleman, C.T.J. Lin, and T. Roberts (1998), "Topics in Microbial Risk
Assessment: Dynamic Flow Tree Process," Risk Analysis, 18(3):309-328.
McCamley, F., and R.K. Rudel (1995), "Graphical Sensitivity Analysis for Generalized
Stochastic Dominance," Journal of Agricultural and Resource Economics, 20(2):403-
403.
McCarthy, M.A., M.A. Burgman, and S. Ferson (1995), Sensitivity Analysis for Models of
Population Viability, Biological Conservation, 73(1):93-100.
McCullagh, P., and J.A. Nelder (1989). Generalized Linear Models. Second Edition. Chapman
and Hall: London.
McElroy, D.M., L.A. Jaykus, and P.M. Foegeding (2000), "Validation and Analysis of Modeled
Predictions of Growth of Bacillus Cereus Spores in Boiled Rice," Journal of Food
Protection, 63(2):268-272.
McKone, T.E. (1996), "Overview of the Risk Analysis Approach and Terminology: the Merging
of Science, Judgment and Values," Food Control, 7(2):69-76.
67
McRae, G.J., J.W. Tilden, and J.H. Seinfeld (1982), Global Sensitivity Analysis A
Computational Implementation of the Fourier Amplitude Sensitivity Test, Computers
and Chemical Engineering, 6(1):15-25.
Mead, P.S., S. Laurence, D. Vance, L.F. McCaig, J.S. Bresee, C. Shapiro, P.M. Griffin, and R.V.
Tauxe (1999), "Food-Related Illness and Death in the United States," Emerging
Infectious Diseases, 5(5):607-625.
Menard, Scott (1995). Applied Logistic Regression Analysis. Sage Publications: Thousand Oaks,
California.
Merz, J.F., M.J. Small, P.S. Fischbeck (1992), Measuring Decision Sensitivity: A Combined
Monte Carlo Logistic Regression Approach, Medical Decision Making, 12(3):189-
196.
Mohanty, S., and T.J. McCartin (1998). Total-System Performance Assessment (TPA) Version
3.2 Code: Module Descriptions and User's Guide. Center for Nuclear Waste Regulatory
Analyses: Southwest Research Institute, Texas.
Montgomery, D.C. (1997). Design and Analysis of Experiments. Wiley and Sons Ltd.: New
York.
Morgan, M.G., and M. Henrion (1990). Uncertainty: A Guide to Dealing With Uncertainty in
Quantitative Risk and Policy Analysis. Cambridge University Press: Cambridge, NY.
Moskowitz, H.R. (1997), A Commercial Application of RSM for Ready-to-Eat Cereal, Food
Quality and Preference, 8(3):191-201.
Myers, R.H., and D.C. Montgomery (1995). Response Surface Methodology: Process and
Product Optimization Using Designed Experiments. Wiley and Sons Ltd.: New York.
68
Neter, J., M.H. Kutner, C.J. Nachtsheim, and W. Wasserman (1996). Applied Linear Statistical
Models. Fourth Edition. McGraw-Hill: Chicago, IL.
Notermans, S., G. Gallhoff, M.H. Zwietering, and G.C. Mead (1995), "The HACCP Concept:
Specification of Criteria Using Quantitative Risk Assessment," Food Microbiology,
12(2):81-90.
NRC (1983). Risk Assessment in the Federal Government: Managing the Process. National
Research Council. National Academy Press: Washington, DC.
Oh, B.H., and I.H. Yang (2000), "Sensitivity Analysis of Time-Dependent Behavior in PSC Box
Girder Bridges," Journal of Structural Engineering-ASCE, 126(2):171-179.
Ozaki, I., F. Kimura, and M. Berz (1995), Higher-Order Sensitivity Analysis of Finite-Element
Method by Automatic Differentiation, Computational Mechanics, 16(4):223-234.
Pet-Armacost, J.J., J. Sepulveda, and M. Sakude (1999), Monte Carlo Sensitivity Analysis of
Unknown Parameters in Hazardous Waste Transportation Risk Assessment, Risk
Analysis, 19(6):1173-1184.
Phillips, A., D. Janies, and W. Wheeler (2000), "Multiple Sequence Alignment in Phylogenetic
Analysis," Molecular Phylogenetics And Evolution, 16(3):317-330.
Rall, L.B (1980). Fundamentals of Numerical Computation: Applications of Software for
Automatic Differentiation in Numerical Computation. Computing Supplement No. 2.
Springer Verlag: Berlin, New York.
Reed, J.A., A.A. Afjeh (2000), "Computational Simulation of Gas Turbines: Part 1 - Foundations
of Component-Based Models," Journal of Engineering for Gas Turbines and Power-
Transactions of the ASME, 122(3):366-376.
69
Rodriguez-Camino, E., and R. Avissar (1998), "Comparison of Three Land-Surface Schemes
with the Fourier Amplitude Sensitivity Test (FAST)," Tellus Series A - Dynamic
Meteorology and Oceanography, 50(3):313-332.
Rose, J.B. (1993). "Waterborne Pathogens: Assessing Health Risks," Health and Environment
Digest, 7(3):1-3.
Rossier, P., D.T. Wade, M. Murphy (2001), "An Initial Investigation of the Reliability of the
Rivermead Extended ADL Index in Patients Presenting with Neurological Impairment,"
Journal of Rehabilitation Medicine, 33(2):61-70.
Salehi, F., S.O. Prasher, S. Amin, A. Madani, S.J. Jebelli, H.S. Ramaswamy, and C. T. Drury
(2000), " Prediction of Annual Nitrate-N Losses in Drain Outflows with Artificial Neural
Networks," Transactions of the ASAE, 43(5):1137-1143.
Saltelli, A., and R. Bolado (1998), An Alternative Way to Compute Fourier Amplitude
Sensitivity Test (FAST), Computational Statistics and Data Analysis, 26(4):445-460.
Saltelli, A., K. Chan, and E.M. Scott. Eds (2000). Sensitivity Analysis. John Wiley and Sons,
Ltd.: West Sussex, England.
Saltelli, A., S. Tarantola, and K. Chan (1999), "A quantitative, Model Independent Method for
Global Sensitivity Analysis of Model Output," Technometrics, 41(1): 39-56.
Searle, S. R (1987). Linear Models for Unbalanced Data. John Wiley and Sons: New York.
Sen, A., and M. Srivastava (1990). Regression Analysis: Theory, Methods, and Applications.
Springer-Verlag: New York.
Seward, S. (2000), "Application of HACCP in Food Service," Irish Journal of Agricultural and
Food Research, 39(2):221-227.
70
Shortencarier M.J., and J.C. Helton (1999). "A FORTRAN 77 Program and User's Guide for the
Statistical Analyses of Scatter plots to Identify Important Factors in Large-Scale
Simluations." SAND99-1058, Prepared by Sandia national Laboratories, Albuquerque,
New Maxico.
Simon, H.A (1996). The Sciences of the Artificial. Third Edition. MIT Press: Cambridge, MA.
Sinicio, R., R.E. Muir, and D.S. Jayas (1997), "Sensitivity Analysis of a Mathematical Model to
Simulate Aeration of Wheat Stored in Brazil," Postharvest Biology and Technology,
11(2):107-122.
Sobell, L.C., S. Maisto, A.M. Cooper, and M.B. Sobell (1982), "Comparison of Two Techniques
to Obtain Retrospective Reports of Drinking Behavior from Alcohol Abusers," Addictive
Behaviors, 7(1):33-38.
Sobol, I.M. (1993), Sensitivity Estimates for Nonlinear Mathematical Models, Mathematical
Modeling and Computation, 1(4):407-414.
Song, F.J., A.M. Glenny, and D.G. Altman (2000), "Indirect Comparison in Evaluating Relative
Efficacy Illustrated by Antimicrobial Prophylaxis in Colorectal Surgery," Controlled
Clinical Trials, 21(5):488-497.
SPLUS (2000). Guide to Statistics, Volume I and II. Data Analysis Product Division, MathSoft
Inc.: Seattle, WA.
Starr, M.K., and C.S. Tapiero (1975), "Linear Break-Even Analysis Under Risk," Operational
Research Quarterly, 26(4):847-856.
Steel, R.G.D., J.H. Torrie, and D.A. Dickey (1997). Principals and Procedures of Statistics A
Biometric Approach. 3
rd
Edition. WCB McGraw-Hill: Boston, Massachusetts.
71
Stiber, N.A., M. Pantazidou, and M.J. Small (1999), Expert System Methodology for
Evaluating Reductive Dechlorination at TCE Sites, Environmental Science and
Technology, 33(17):3012-3020.
Taylor, B.C., J.L. Stott, M.A. Thurmond, and J.P Picanso (1992), "Alteration in Lymphocyte
Subpopulations in Bovine Leukosis Virus-Infected Cattle," Veterinary Immunology and
Immunopathology, 31(1-2):35-47.
Tienungoon, S., D.A. Ratkowsky, T.A. McMeekin, and T. Ross (2000), "Growth Limits of
Listeria Monocytogenes As a Function of Temperature, pH, NaCl, and Lactic Acid,"
Applied And Environmental Microbiology, 66(11):4979-4997.
Vidmar, T.J., and J.W. McKean (1996), A Monte Carlo Study of Robust and Least Squares
Response Surface Methods, Journal of Statistical Computation and Simulation, 54(1):1-
18.
von Winterfeldt D, and Edwards W (1986) Decision Analysis and Behavioral Research.
Cambridge, UK: Cambridge University Press, 1986: 399-405.
Vose, David (2000). Risk Analysis: A Quantitative Guide. 2
nd
Edition. John Willey & Sons, Ltd:
New York.
Walpole, R.E., and R.H. Myers (1993). Probability and Statistics for Engineers and Scientists.
5
th
Edition. Macmillan Publishing Company: New York.
Ward, M.P., and T.E. Carpenter (1996), "Simulation Modeling of the Effect of Climatic Factors
on Bluetongue Virus Infection in Australian Cattle Herds - 1. Model Formulation,
Verification and Validation," Preventive Veterinary Medicine, 27(1-2):1-12.
72
Weiss, R (1996), "An Approach to Bayesian Sensitivity Analysis," Journal of the Royal
Statistical Society Series B-Methodological, 58(4):739-750.
WHO (2001), "Microbial Risk Assessment in Food," World Health Organization.
http://www.who.int/fsf/mbriskassess, Viewed: March, 21, 2001.
Williams, S., K. Varahramyan, and W. Maszara (1999), Statistical Optimization and
Manufacturing Sensitivity Analysis of 0.18 m SOI MOSFETs, Microelectronic
Engineering, 49(3-4):245-261.
Winter, B.J., D.R. Brown, and K.M. Michels (1991). Statistical Principles in Experimental
Design. Third Edition. McGraw-Hill Inc.: New York.
Wotawa, G., A. Stohl, and H.KrompKolb (1997), "Estimating the Uncertainty of a Lagrangian
Photochemical Air Quality Simulation Model Caused by Inexact Meteorological Input
Data," Reliability Engineering & System Safety, 57(1):31-40.
Zwietering, M.H., and S.J.C. van Gerwen (2000), "Sensitivity Analysis in Quantitative Microbial
Risk Assessment," International Journal of Food Microbiology, 58(3):213-221.

73

Figure 1. Example of chart for representing difference in log odds ratio.
(Source: Stiber et al., 1999)

Figure 2. Example of break-even analysis.

- 1
- 0. 5
0
0. 5
1
1. 5
2
L
o
g
-
o
d
d

R
a
t
i
o
T
e
r
m
i
n
a
l

E
l
e
c
t
r
o
n
H
y
d
r
o
g
e
n
R
e
d
o
x

P
o
t
e
n
t
i
a
l
D
i
s
s
o
l
v
e
d

O
r
g
.

C
B
T
E
X
T
P
H
O
x
y
g
e
n
T
e
m
p
e
r
a
t
u
r
e
p
H
D
C
E
V
i
n
y
l
e

C
h
l
o
r
i
d
e
E
h
e
n
e

&

E
t
h
a
n
e
M
e
t
h
a
n
e
C
h
l
o
r
i
d
e

-50
-30
-10
10
30
50
70
90
0 0.2 0.4 0.6 0.8 1
P of Success of Operation
U
t
i
l
i
t
y

o
f

M
e
d
i
c
a
t
i
o
n
,

M
Medication is favored
Operation is favored
Indifference or
break-even line
74
Table I. Forward stepwise regression analysis for mean population dose.
(Source: Helton et al., 1995)
Steps Inputs Standardized Regression Coefficients R
2

1 PSMCS134 0.62 0.40
2 TFMCS 0.27 0.78
3 PSMI131 0.32 0.86
4 PSMCS137 0.19 0.89
5 TFMI131 0.13 0.90
6 IF1 0.11 0.92
7 AF - 0.10 0.92
8 TFMSR 0.09 0.93
9 FRCTCMP 0.09 0.94
10 PSMSR90 0.09 0.94

75
Table II. Overview of Comparison of the Methods
Methods Applicability Computational Issues Representation of Sensitivity Best Use of Methods
Nominal
Range
Sensitivity
Analysis
Deterministic
model.
Need nominal range for each input,
potentially time consuming.
Ratios, percentages. Does not include
effect of interactions or correlated
inputs. Easy to understand.
Key inputs for linear
models, verification
and validation.
LOR Deterministic
model with output
as a probability.
Need nominal range for each input,
Ratios, percentages. Does not include
effect of interactions or correlated
inputs. Easy to understand.
Key inputs for linear
models, verification
and validation.
Breakeven
Analysis
Models used to
choose among
alternatives.
Complex for model with many decision
options and/or more than two inputs,
Graphical representation. Robustness of
solution.
AD Locally different-
tiable models.
Require specific software (e.g.,
ADIFOR)
Local sensitivity measures, such as
sensitivity coefficients
Potential key inputs,
verification.
Regression To results from
probabilistic
simulation
Must specify functional form,
computation time and value of solution
depends on specific techniques used.
R
2
, t-ratios for regression coefficients,
standard regression coefficients, and
others.
key inputs, joint
effect of multiple
inputs, verification.
ANOVA
(MF)
a
Probabilistic
models
Time consuming for a large number of
inputs with interactions.
F-value, Tukey test coefficients, and
other that are calculated at different
stages of ANOVA.
key inputs, joint
effect of multiple
inputs, verification.
RSM Any deterministic
model
Developed using a variety of techniques,
some require functional forms, others do
not; may require extensive runs to
generate a calibration data set.
Graphical, evaluation of functional
form, method-dependent measures.

Model of models,
used with other SA
methods to save time.
FAST
(MF)
Probabilistic
models
Better with no interactions/higher order
input. Caution against discrete inputs.
Portion of output variance attributable
to each input
Key inputs including
combined effect,
verification.
MII
(MF)
Probabilistic
model.
Complex, no computer code available,
Time consuming.
Amount of "mutual information" about
the output provided my each input, also
graphs of intermediate stages.
Key inputs including
combined effect.
Scatter
Plots
(MF)
Probabilistic
model.
Easy, time requirement depends on the
number of input/outputs.
Graphical, no quantitative sensitivity. Verification and
validation.
a
MF = denotes a model-independent approach

Frey

Uploaded by

Copyright:

Available Formats

Frey

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Frey

Uploaded by

Copyright:

Available Formats

1

IDENTIFICATION AND REVIEW OF SENSITIVITY ANALYSIS METHODS

= error for the i

You might also like