Dealing with uncertainty in ecosystem models: lessons from a complex salmon model

Ecosystem models have been developed for assessment and management M a wide variety of environments. As model complexity increases, it becomes more difficult to trace how imperfect knowledge of internal model parameters, data inputs, or relationships among parameters might impact model results, affecting predictions and subsequent management decisions. Sensitivity analysis is an essential component of model evaluation, particularly when models are used to make management decisions. Results should be expressed as probabilities and should realistically account for uncertainty. When models are particularly complex, this can be difficult to do and to present in ways that do not obfuscate essential results. We conducted a sensitivity analysis of the Ecosystem Diagnosis and Treatment (EDT) model, which predicts salmon productivity and capacity as a function of ecosystem conditions. We used a novel "structured sensitivity analysis" approach that is particularly useful for very complex models or those with an abundance of interconnected parameters. We identified small, medium, and large plausible ranges for both input data and model parameters. Using a Monte Carlo approach, we explored the variation in output. prediction intervals, and sensitivity indices, given these plausible input distributions. The analyses indicated that, as a consequence of internal parameter uncertainty, EDT productivity and capacity predictions lack the precision needed for many management applications. However, EDT prioritization of reaches for preservation or restoration was more robust to given input uncertainties, indicating that EDT may be more useful as a relative measure of fish performance than as an absolute measure. Like all large models, if EDT output is to be used as input to other models or management tools it is important to explicitly incorporate the uncertainty and sensitivity analyses into such secondary analyses. Sensitivity analyses should become standard operating procedure for evaluation of ecosystem models.