Comparison of model-averaging and single-distribution approaches to estimating species sensitivity distributions and hazardous concentrations for 5% of species.

Abstract

Estimation of species sensitivity distributions (SSDs) and hazardous concentrations for 5% of species (HC5s) by fitting a statistical distribution to toxicity data for multiple species is essential in ecological risk assessment of chemicals. Given the challenge of selecting the appropriate statistical distribution in SSD estimation, a model-averaging approach that involves fitting multiple statistical distributions and using weighted estimates to derive HC5s is appealing. However, the effectiveness of this approach compared with SSDs based on a single statistical distribution (i.e., single-distribution approach) has not been thoroughly examined. We aimed to compare the model-averaging approach with the single-distribution approach based on log-normal, log-logistic, Burr type III, Weibull, and gamma distributions to estimate HC5s. For this comparison, we selected 35 chemicals with available toxicity data for more than 50 species, enabling the direct calculation of reference HC5 values from the 5th percentiles of the toxicity distributions. For each chemical, we examined the deviations between the reference HC5 value and HC5 estimates derived from SSDs based on toxicity data for 5-15 species subsampled from the complete dataset using model-averaging and single-distribution approaches. This subsampling simulated the typical limitations of available toxicity data. The deviations observed with the model-averaging approach were comparable with those from the single-distribution approach based on the log-normal, log-logistic, and Burr type III distributions. Although use of specific distributions often resulted in overly conservative HC5 or HC1 estimates, our results suggest that the precision of HC5/HC1 estimates would not substantially differ between the model-averaging approach and the single-distribution approach based on log-normal and log-logistic distributions. We further discuss the circumstances under which model-averaging and single-distribution approaches are better suited for estimating HC5s.