UW Aquatic & Fishery Sciences Quantitative Seminar

Ashley Steel

NOAA Fisheries, Northwest Fisheries Science Center

Sensitivity analysis of a model used for fisheries management: Making sense of 10,000 parameters

Authors: Paul McElhany and Ashley Steel, NOAA Fisheries

Abstract

Ecosystem models have been developed for a number of environments and management applications such as endangered species management. As model complexity increases, it becomes more difficult to trace the path from imperfect knowledge of internal model parameters and data inputs to modeled predictions and management decisions. We provide a case study of the sensitivity analysis of a large and complex fish-habitat model, Ecosystem Diagnosis and Treatment (EDT) and present three sensitivity analyses of the model conducted by three different government agencies for three specific management purposes. We describe in detail a novel “structured sensitivity analysis” approach that is particularly useful for very complex models. In this structured analysis, we identified small, medium and large plausible ranges for all 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. Finally, we draw combined conclusions from the three analyses. The details of how each agency conducted and utilized sensitivity analyses are outlined; trade-offs between simpler and more intensive sensitivity analyses are described. Combined insights on the EDT model include identification of input parameters to which the model is surprisingly insensitive and quantification of prediction intervals. We conclude that known uncertainties in input data and internal parameters lead to large prediction intervals around estimates of population abundance and productivity but that identification of high priority reaches for restoration and preservation is relatively robust to these known uncertainties. We recommend that sensitivity analyses be considered an essential component of applying any model to make management decisions.