Environmental Engineering Reference
In-Depth Information
0.5
Fig. 14.9 Three different
statistical models (logistic,
normal and triangular) can fit
data (not shown) equally
well, but result in dissimilar
estimates of numerical
outcomes (in the example: the
5th percentile cut-off of
impacts would result in three
different numerical values of
the HC5. Note that the X-axis
represents a log-scale, so that
the HC5-values are all near
0.01mg/kg dw .Thethree
horizontal bars indicate that
the Confidence Intervals of all
three HC5-estimates overlap
Logistic
0.3
0.2
Normal
Triangular
0.1
5%
-3
-2
-1
0
1
2
3
Three numerical estimates of HC5
values of HC5 would lead to three different proposals for a soil quality standard,
while only one could be published as a formally adopted value. The model choice
thus always matters in terms of societal and practical impacts, but much less so
statistically.
To give an impression of the possible practical impacts of choosing alterna-
tive statistical approaches for the same input data, as in Fig. 14.9 , we refer to
Verdonck et al. ( 2001 ). These authors concluded on the basis of statistical com-
parisons between alternative SSD-modeling approaches that those approaches vary
in statistical robustness per se , that they differ in their sensitivity for decreasing
sample sizes, and that the numerical estimates for HC5s as generated may differ by
a factor of 5. Most importantly, they said there is no apriori reason to prefer any
of the possible statistical approaches. These findings illustrate three relevant issues.
First, the factor of 5 mentioned by Verdonck is much smaller than the uncertainty
factors (of 10 and 100) that are often applied in Effect Assessments, according to
various guidance documents, to derive a quality standard from (e.g.,) the lowest test
NOEC. Such uncertainty factors are often 10, 100 or 1000. Second, though the point
estimates may vary by a factor of 5, there may be overlap of the confidence intervals
of the estimates generated by different methods (see Fig. 14.9 ). Third, based on val-
idation studies, on the concentration scale of a field gradient all HC5s may be in the
range of “no statistically significant or visible signal of response for the endpoint
of concern” (compare neutral response concentration range in Fig. 14.2 ). Whether
these three issues apply in other cases could be a subject of study, e.g., by fitting
alternative models to the data, and by comparison of the results amongst each other
and with results from an uncertainty factor approach. Next to these statistical explo-
rations, it is generally undoable to find compound-related case studies for checking
the validation issue.
Because a factor of 5 may make a significant difference in the regula-
tion of contaminants (Y
X) and in the land area requiring remediation, it is
Search WWH ::




Custom Search