|
(064) Comparing statistical techniques for uncertainty assessment of species sensitivity distributions: effect of sample size. Verdonck, Frederik1, Jaworska, Joanna2, Janssen, Colin3, Vanrolleghem, Peter1, 1 2 3 ABSTRACT- Species sensitivity distributions (SSD) of laboratory derived ecotoxicity data are commonly used in environmental risk assessments to derive 'safe concentrations' of contaminants. For example, in EU environmental risk assessment practices, the 5th-percentile of the SSD (HC5: hazardous concentration) is used to set these environmental quality standards. Although, several techniques are available to characterize uncertainty of a SSD and HC5, many of these give different results. The choice of the underlying (statistical) distribution has a large effect. Additionally, once a technique has been selected, the question arises how many species are needed to calculate HC5 with a desired accuracy. A comparison is made between maximum likelihood estimation (assuming lognormal and loglogistic distribution), Bayesian approaches (assuming lognormal distribution) and nonparametric bootstrapping (using Hazen plotting and the interpolated empirical distribution function) techniques to characterise uncertainty and variability using small ecotoxicity data sets. Accurateness depends on an appropriate sample size (for determining a SSD). If the data are lognormal or loglogistic distributed and when the parametric techniques are used, HC5 confidence intervals are accurate, even for very small sample size (five data points already appeared to be sufficient). However, these techniques are very sensitive to outliers. Using nonparametric techniques, the HC5 confidence intervals are overestimated, but they do not depend on an assumption of an underlying distribution and are less sensitive to outliers. Ten data points were sufficient for the simulated data sets when using the nonparametric bootstrap with the interpolated empirical distribution function and the Hazen plotting system. The importance of selection the appropriate statistical SSD and the minimum number of taxa represented in the SSD will be discussed in the context of current scientific and regulatory practices. Key words: uncertainty, confidence intervals, species sensitivity distribution, minimum/optimal sample size |
