Chemical Hazard Assessment under Uncertainty Evaluated by Robust Bayesian Evidence Synthesis

Abstract: Chemical hazard assessment is a procedure in chemical regulation to set which concentrations of chemical substances that are safe to an ecological system, i.e. where a majority of the species in the systems is not negatively affected by the substance. The Species Sensitivity Distribution (SSD) is an assessment model to estimate hazardous concentration based on toxicity data from representative species from an ecological system. In its current use, the hazard assessment using SSD approach is only to be applied based on enough large sample size of toxicity data, which in absence of large sample sizes leads to overly conservative regulation, and the choice of hazardous concentration is based on statistical properties of the SSD, without any explicit loss function. In this thesis, the hazard assessment using SSD was implemented as a Bayesian evidence synthesis integrating different sources of evidence to calibrate the assessment model and propagate uncertainty to a decision model. The hazardous concentration was then the Bayes optimal decision that minimizes the undesired effects caused by the chemical substance. Loss was determined by a linear-exponential loss function, which captures conservatism in an effective way and is useful to find thresholds. The aim of this thesis was to evaluate the impact of uncertainty on the hazardous concentration due to small sample size, measurement errors in toxicity data and, since uncertainty was quantified by Bayesian probabilities, the choice of priors for the SSD as well. The impact of uncertainty on the hazard concentration was evaluated by a simulation study. It was shown that choice of prior on the mean and spread of the SSD matters when sample size was less than 30 toxicity data. Quality in data, manifested as estimation errors of toxicity values, had a profound influence on choosing hazardous concentration. When the data was sufficiently large, choice of the priors and the quality of the data almost had no impact on choosing the optimal hazardous concentration. It was suggested that a higher value on a certain parameter for the loss function might compensate for the small data or data with poor quality. Several repetitions of simulation are needed to make result more reliable.