Uncertainty analysis methods

for human or ecological health might not be considered. Because

scenario uncertainty is conceptual, it is difficult to quantify.

b. Model uncertainty arises from the necessary simplification of real-world

relationships among environmental components. For example,

bioaccumulation models are subject to imperfect knowledge about the

potential for accumulation of contaminants by benthic invertebrates. This

source of uncertainty is quantifiable, but the required level of effort

necessary to describe the uncertainty could be substantial. It could be

partially quantified by comparing model predictions with field

measurements. In addition, results could be compared among models that

have different structures, but are designed to predict the same output.

c. Parameter uncertainty arises from lack of knowledge about the true

distribution of model parameters, perhaps due to measurement error.

Measurement error is a common example of parameter uncertainty. This

source of uncertainty can be measured in various ways. For example

collection and analysis of field sample duplicates can provide an estimate

of uncertainty that can be reduced by improving sampling and analytical

methods. Of all the sources of uncertainty, parameter uncertainty is

probably the easiest to quantify.

Uncertainty analysis methods

Uncertainty analysis methods range from qualitative approaches (e.g., expert

judgment) to quantitative approaches, such as descriptive statistics and two-

dimensional Monte Carlo analysis. The choice among these methods depends on

project objectives and the quantity and quality of available data. Several available

methods are described as follows:

a. Descriptive statistics such as confidence intervals and standard deviation

estimates can provide measures of uncertainty and variability.

b. Probabilistic analysis (with or without Latin Hypercube simulation)

(Vose 1996; Burmaster and von Stackelberg 1991) is used to describe

potential outcomes in terms of probability. With probabilistic

approaches, one can calculate correlation between model input

distributions and the predicted output distribution (e.g., exposure or risk)

to identify inputs that strongly influence predictions. One probabilistic

approach, two-dimensional Monte Carlo (Frey 1992; Burmaster and

Wilson 1996), allows the conceptual differences between variability and

uncertainty to be assessed separately. This separation facilitates the

application of model results to policy questions about exposed

populations and ecosystems, exposure levels, and research needs (Bogen

and Spear 1987; Frey 1992).

c. Value of Information (VOI) (Thompson and Evans 1997) is a decision

analytic framework that is an extension of uncertainty analysis. VOI is

used to determine whether the cost of obtaining information prior to

choosing a course of action is justified given the benefit of having that

information as input to the decision.

Chapter 1 Introduction