5.1 Sensitivity Testing and Uncertainty Handling
89
Chap. 4 “Testing and Validating Against Historic Spills”). In such a case of applying
a complex model and multitude of uncertain parameters, it is important to realize and
accept that we do not know the “true risk” as a number as such but need to ensure
that the model does not underestimate the risk and can be used to compare risks.
5.2 Methods Used in Sensitivity Testing
To test the sensitivity of the calculations towards numerical variation in the input
parameters both deterministic and stochastic tests were carried out.
5.2.1 Deterministic Testing
In the deterministic testing, the impact calculations, lag- and restoration calculations within the model and its sub-models were tested by breaking them into the
individual functions. By holding the other parameters constant, the input parameters were varied one by one and the resulting endpoints calculated. The results are
available as graphs. The reader is encouraged to read the full reports with method
description and results in: Bjørgesæter and Damsgaard-Jensen (2018) and Stephansen
and Bjørgesæter (2017).
These simpler deterministic tests holding one parameter fixed at a time (One-At
A-Time tests, OAT) are useful to study the direct output of varying single parameters,
and thus get better acquainted with the results of the individual calculations. However,
these deterministic tests are unsuitable for handling the many dimensions of variation
of the input parameters, for which the global stochastic sensitivity methods are used
(Marino et al. 2008).
The range in parameter values found in the literature studies during methodology
development was used to define the range between the minimum and maximum
values but these ranges were not used to limit the sensitivity analyses performed
in the next step (see the references in the methodology development, Chap. 3 and
references to Tables A.1, A.2 and A.3 in Supplementary Information 1).
A deterministic approach requires few simulations and is therefore valuable for
examining models that may become costly in terms of computer time (e.g. testing
oil drifts statistics used in the models).
The disadvantages of the deterministic approach include; only a few discrete
outcomes are considered, it gives equal weight to each outcome, and possible interdependence between inputs are difficult to identify and quantify. Assessing the likelihood of different outcomes is therefore not possible with deterministic testing, and
it is difficult to identify and rank the input parameter in terms of importance on the
model output.
89
Chap. 4 “Testing and Validating Against Historic Spills”). In such a case of applying
a complex model and multitude of uncertain parameters, it is important to realize and
accept that we do not know the “true risk” as a number as such but need to ensure
that the model does not underestimate the risk and can be used to compare risks.
5.2 Methods Used in Sensitivity Testing
To test the sensitivity of the calculations towards numerical variation in the input
parameters both deterministic and stochastic tests were carried out.
5.2.1 Deterministic Testing
In the deterministic testing, the impact calculations, lag- and restoration calculations within the model and its sub-models were tested by breaking them into the
individual functions. By holding the other parameters constant, the input parameters were varied one by one and the resulting endpoints calculated. The results are
available as graphs. The reader is encouraged to read the full reports with method
description and results in: Bjørgesæter and Damsgaard-Jensen (2018) and Stephansen
and Bjørgesæter (2017).
These simpler deterministic tests holding one parameter fixed at a time (One-At
A-Time tests, OAT) are useful to study the direct output of varying single parameters,
and thus get better acquainted with the results of the individual calculations. However,
these deterministic tests are unsuitable for handling the many dimensions of variation
of the input parameters, for which the global stochastic sensitivity methods are used
(Marino et al. 2008).
The range in parameter values found in the literature studies during methodology
development was used to define the range between the minimum and maximum
values but these ranges were not used to limit the sensitivity analyses performed
in the next step (see the references in the methodology development, Chap. 3 and
references to Tables A.1, A.2 and A.3 in Supplementary Information 1).
A deterministic approach requires few simulations and is therefore valuable for
examining models that may become costly in terms of computer time (e.g. testing
oil drifts statistics used in the models).
The disadvantages of the deterministic approach include; only a few discrete
outcomes are considered, it gives equal weight to each outcome, and possible interdependence between inputs are difficult to identify and quantify. Assessing the likelihood of different outcomes is therefore not possible with deterministic testing, and
it is difficult to identify and rank the input parameter in terms of importance on the
model output.