The sensitivity of multiple output statistics to input parameters in a krill–predator–fishery ecosystem dynamics model

There is a global need to develop strategic frameworks for assessing uncertainty in ecosystem dynamics models. Such models have been used within CCAMLR to evaluate options for managing the Antarctic krill fishery in the Scotia Sea and southern Drake Passage. The model analysed here required 2 311 input values for each of four scenarios and produced 68 output statistics. Small perturbations to input values affected output statistics indicating the status of predator groups more than they affected statistics indicating the status of the target stock or the fishery. Output statistics were most sensitive to a parameter controlling predator recruitment through pre-recruit mortality. A parameter mediating the effect of a forcing function on krill recruitment, which was used to condition the model on past dynamics, was also important, and some of the parameter estimates resulting from conditioning were unstable. This highlights the tension between the parameter stability benefits of well-constrained models and the use of model conditioning to identify plausible alternative hypotheses in data-poor situations. Apparent sensitivity is a function of both input values and output statistics. Clearer specification of ecosystem-based management objectives would help to identify the important statistics for consideration when assessing uncertainty in ecosystem dynamics models.