O'Hagan, A, Crane, M., Grist, E. P. M. and Whitehouse, P.
Department of Probability and Statistics, University of Sheffield, Sheffield, England, Crane Consultants, Faringdon, Oxfordshire, England, CSIRO Marine Research, Hobart, Tasmania, Australia and Environment Agency, Wallingford, Oxfordshire, England
Publication details: Submitted to Applied Statistics.
A species sensitivity distribution (SSD) is the probability distribution of some measure of toxicity to a certain chemical in a population of animal species. Given data consisting of estimates of toxicity for a number of species present in the habitat of interest, the SSD is typically estimated by assuming that these values are a random sample from a lognormal distribution and estimating the lognormal parameters. The principal deficiency of this approach is the assumption that the data are from a random sample of species. In practice, the species for which data are available are determined in non-random ways and are likely to be highly non-representative of the population.
We present a method of inference about SSDs that draws on expert judgement about which species are likely to be more sensitive to the chosen chemical. The expert judgements allow us to take some account of non-representativeness of the available data. We adopt a hierarchical random-effects model which recognises that species within the same family are likely to have similar sensitivities, and employ a Bayesian approach to analysis that allows direct inference about quantiles of the SSD.
Keywords: Bayesian inference; censored data; chlorpyrifos; environmental standards; HC5; hierarchical model; LC50; SSD; toxicology.