Bayesian hierarchical modeling of oil toxicity in fish early life stages: Supporting oil-spill risk assessments in data-scarce situations.

The retrospective nature of oil-spill impact assessments has led to uneven data availability across fish species and oil types, increasing uncertainty in parameter estimates and limiting the reliability of toxicity models. To address this, we applied a suite of Bayesian hierarchical dose-response models to a dataset of oil-induced fish early life stage-mortality. Bayesian hierarchical models enable information borrowing across groups thereby improving estimation in data-scarce contexts. Our results demonstrated that our models reduced posterior uncertainty compared with non-hierarchical models and improved the generalizability of model estimates by smoothing out sampling variation due to small sample size. Despite substantial data imbalance across species and oil types, the models showed strong predictive performance, suggesting that this approach can reduce the need for costly new data generation. The Bayesian framework is transparent about uncertainty and thus considers differing risk attitudes of decision makers. It also enables Bayesian updating that can be utilized to update parameter estimates and model structures efficiently with new information and provides, together with information borrowing, a robust foundation for machine-learning applications in data-scarce settings. By integrating parameters for biological and environmental variability, our models enhance the capacity to evaluate diverse oil-spill scenarios within fish population dynamics models and support more accurate, cost-effective risk assessments and decision-making in data-scarce settings. In addition to oil spill risk assessment, the framework presented here is transferable to other ecotoxicological risk assessment contexts as well.