Environmental Engineering Reference
In-Depth Information
14.8.1.3 Example Data Bases
Various data bases have been made to facilitate the collection of input data for SSDs.
Examples of compilations of all kinds of ecotoxicity data are the U.S. EPA's Ecotox
database (U.S.EPA 2002 ) and the Dutch RIVM e-toxBase (Wintersen et al. 2004 ).
Many researchers have their own data collections.
14.8.2 Ingredient 2: The Statistical Approach
14.8.2.1 Options for Model Choice
Figure 14.3 shows that one can use raw sensitivity data directly (bar diagram), and
that one can use models, like parametric curve fitting, or more complex approaches,
like nonparametric or Bayesian regression.
Frequently, SSDs have been derived using a log-logistic or a log-normal model.
Alternatively, multi-modality of the SSD can be considered when the bell-shaped
curve has two peaks, see e.g., Aldenberg and Jaworska ( 1999 ). In such cases, there
are two subsets of ecotoxicity data. A specifically sensitive subgroup can be distin-
guished, such as insects in the case of insecticides. Non-insects can then be modeled
with a separate Normal distribution shifted to the insensitive side. In this case, the
user may want to make two SSDs, one for each subgroup of species, instead of one
bimodal model.
The user may wish to give different weights to the different input data (instead
of weighting all data equally; see, e.g., Duboudin et al. ( 2004 )), because some
of the species may be considered more important or relevant for the soil health.
Earthworms, for example, are sometimes considered to be “ecosystem engineers”
(see, e.g., Jouquet et al. ( 2006 )), due to their large capacities in soil turnover and
aeration, and arguably they might be given a higher “weight” in deriving the SSD
for a contaminant. How to give those weights is an issue of expert judgment, and is
not specific to SSD-modeling.
14.8.2.2 Selecting a Model
All statistical approaches have advantages and disadvantages. Not all of these issues
are treated here. Posthuma et al. ( 2002b ) provide references to the different models
and approaches. For practical users it is not necessary to fully understand the statis-
tical refinements of each of the methods, although all mathematical approaches and
practice details differ in kind. Three theoretical issues should be considered in the
model choice:
1. the fit of the chosen model to the data (misfits may identify specific aspects of
the input data);
2. despite an overall good fit, the possible bias of the model in the concentration
range of interest, and finally
3.
the validity of the output for the case of concern
Search WWH ::




Custom Search