6. Likelihood of Introducing Nonindigenous Organisms
87
domestic regulation of wheat shipments (USDA 1996a–d, 1997) to limit the
spread of Karnal Bunt, each of the 24 scenarios represented a different program
option. We construct the probabilistic risk assessment models as fault models.
Each independent event (primary node) in the model represents a “failure” with
respect to risk management (the probabilities of the subnodes may be additive,
conditional, and so on). The final estimate of risk depends on the mathematical
relationship among the nodes. In developing the risk assessment model, it is
important to conduct a units analysis to ensure that the model inputs, together with
the mathematical relationship among the nodes, result in the appropriate output.
The scenario analysis provides a conceptual framework for estimating risk.
Estimation of Input Values: Choosing Distributions and
Distribution Parameters to Represent Estimated
Frequencies and Probabilities
Estimates of risk are only as precise and accurate as the estimates used as input
values. We use the best scientific information available for assessments. Ideally,
existing data would provide the basis for direct estimation of model inputs. However, scientific experiments are seldom conducted specifically to provide these
estimates for risk assessments, and results are seldom provided that can be used
directly in our models. In addition, because most of our risk assessments are
conducted to support decisions that must be made within relatively narrow time
frames, research programs can seldom be conducted to provide data specifically
for our assessments. Fortunately, this situation has already started to change as
risk-based decisions for trade in agricultural commodities become the international standard. Agricultural research has already started in support of this type of
risk assessment.
For now, however, a variety of biological data usually is available that is
pertinent to the needed probability. These data are reviewed, and professional
judgment is used to represent the available data regardless of whether estimates
are characterized as point estimates or distributions of possible values. We base
estimates on pest interception records, the known biology of the organism being
assessed (or the known biology of related taxa), expert judgment based on laboratory experience with the pest or related organisms, expert judgment based on field
experience with the pest or related organisms, expert judgment based on experience conducting commodity inspections at ports of entry or in the exporting
country, and experience working with export programs and export-quality
commodities.
Probabilistic risk assessments are based on Monte Carlo simulations. Input
values are characterized as probability density functions (PDFs) of possible
values. Normal and lognormal distributions are familiar examples of PDFs. Using
PDFs provides an explicit and transparent method to account for the uncertainty
inherent in estimated probabilities for events. The two basic components of a PDF
are the shape of the distribution and the values of distribution parameters such as
87
domestic regulation of wheat shipments (USDA 1996a–d, 1997) to limit the
spread of Karnal Bunt, each of the 24 scenarios represented a different program
option. We construct the probabilistic risk assessment models as fault models.
Each independent event (primary node) in the model represents a “failure” with
respect to risk management (the probabilities of the subnodes may be additive,
conditional, and so on). The final estimate of risk depends on the mathematical
relationship among the nodes. In developing the risk assessment model, it is
important to conduct a units analysis to ensure that the model inputs, together with
the mathematical relationship among the nodes, result in the appropriate output.
The scenario analysis provides a conceptual framework for estimating risk.
Estimation of Input Values: Choosing Distributions and
Distribution Parameters to Represent Estimated
Frequencies and Probabilities
Estimates of risk are only as precise and accurate as the estimates used as input
values. We use the best scientific information available for assessments. Ideally,
existing data would provide the basis for direct estimation of model inputs. However, scientific experiments are seldom conducted specifically to provide these
estimates for risk assessments, and results are seldom provided that can be used
directly in our models. In addition, because most of our risk assessments are
conducted to support decisions that must be made within relatively narrow time
frames, research programs can seldom be conducted to provide data specifically
for our assessments. Fortunately, this situation has already started to change as
risk-based decisions for trade in agricultural commodities become the international standard. Agricultural research has already started in support of this type of
risk assessment.
For now, however, a variety of biological data usually is available that is
pertinent to the needed probability. These data are reviewed, and professional
judgment is used to represent the available data regardless of whether estimates
are characterized as point estimates or distributions of possible values. We base
estimates on pest interception records, the known biology of the organism being
assessed (or the known biology of related taxa), expert judgment based on laboratory experience with the pest or related organisms, expert judgment based on field
experience with the pest or related organisms, expert judgment based on experience conducting commodity inspections at ports of entry or in the exporting
country, and experience working with export programs and export-quality
commodities.
Probabilistic risk assessments are based on Monte Carlo simulations. Input
values are characterized as probability density functions (PDFs) of possible
values. Normal and lognormal distributions are familiar examples of PDFs. Using
PDFs provides an explicit and transparent method to account for the uncertainty
inherent in estimated probabilities for events. The two basic components of a PDF
are the shape of the distribution and the values of distribution parameters such as