Ecological Risk Assessment of Diffuse and Local Soil Contamination Using Species Sensitivity Distributions - Dealing with Contaminated Sites: From Theory Towards Practical Application - page 651

approach, one can check whether the model fits the data (or not), by applying
goodness-of-fit testing, and set apriori fit criteria to accept or reject a fitted model.
A software program like ETX supports this kind of statistical evaluation (see Van
Vlaardingen et al. 2004 ). Models that show a bad fit to the data can thus be rejected
on a chosen statistical criterion, irrespective of the number of data and the set of
test species.
Second, the input data should be selected in view of the problem definition. For
example, for judging soils at low pH, one could select all test data with a pH near
the pH of the studied soil.
It should be noted that rejecting input data by any data selection criterion has a
trade-off. Low data numbers imply wider confidence intervals (see e.g. Fig. 14.3 ),
and sometimes rejection of an SSD model, implying that another model needs to be
chosen. The past selection of NOECs for the derivation of soil quality standards is,
in terms of this tradeoff, a surprising one. Usually there is much more acute toxicity
data, so that sensitivity distributions can be derived with less (statistical) uncertainty
based on such data. If acute data had been chosen for setting protective standards
(such as that explored by Kooijman ( 1987 )) this would have required post hoc acute
to chronic extrapolation. The acceptability of a particular criterion for data selection
for an assessment depends on the case.

14.9.2 Presenting Confidence Intervals

SSDs, like many statistical models (with large or small sets of input data) can pro-
duce both point estimates and confidence intervals of those estimates. Confidence
intervals represent which values the “true” value could have, given the variability of
the available input data. Beware, that this only pertains to the statistical meaning of
confidence (not to extrapolation in the case of ignorance of e.g. the sensitivities of
the field species).
An example of a statistical analysis of confidence intervals in SSD-outputs is
provided in Fig. 14.8 , for the statistical confidence interval of an HC5. When the
SSD-output is used for the derivation of a formal soil quality standard, the statistical
analysis is followed by a policy choice, because a standard cannot be uncertain; a
formal soil quality standard cannot have the form “a maximum tolerable concentra-
tion of 5 plus or minus 2.5mg/kg dw soil”. Hence, the regulatory context requires a
choice, either for the p in HC p , or (when more precaution is taken) for a lower con-
fidence bound on the HC p (e.g., resulting in a value of 5 minus 2.5 = 2.5mg/kg dw

in the example shown above).
In a similar way, one can quantify the confidence interval for estimates of
PAF (Y), given an actual soil concentration (X). In this case, it is possible to
present a comparison between two contaminated sites, by demonstrating for exam-
ple that site A would induce impacts for 50% of the species (confidence interval
45-55%), and site B for 80% of the species (75-85%). Given those small and non-
overlapping confidence intervals, site B clearly presents a greater risk for the test set
(and probably also for field species).