Workshop on the review and future of state space stock assessment models in ICES (WKRFSAM)

The Workshop on the Review and Future of State Space Stock Assessment Models in ICES focused on future directions of state-space assessment models for ICES stocks (WKRFSAM), utilising recent advances in fisheries modelling research to help define best practises. State-space models consist of a process model for unobserved quantities (e.g. true stock abundances) and an observation model for observed quantities (e.g. catches or survey data).

Standard statistical asassessment models do not include stochastic population processes. Prediction is a natural part of the state-space model formulation which is a more practical advantage of the approach for stock assessment. Lognormal observation error models for survey indices can be parameterized such that the mean or median is proportional to the true stock abundance, but this makes little difference on assessment results.

Empirical results for 14 stocks indicated that this is also the case for catch observation models. There is no practical difference whether the mean or median of fishing mortality rates (F’s) are constant for the data period, but there are important differences if F is projected into the future. Including random deviations in the natural mortality rate (M) leads to convergence issues for some stocks, and similar assessment results for other stocks, but this depends on the details of how this is implemented.

Including time- and age-varying M increases model flexibility and potential confounding of F, M, and stock size index catchability, Q. When model parameters are confounded then we can anticipate less robustness. Including external variances and correlations that are reliably estimated will be more relevant and useful in situations where these differ substantially over time, such as when a survey in some year has a large set, or poor coverage, etc.

Two general criteria to select between alternative models are Goodness-of-it and Out-of-sample prediction. Minimizing out-of-sample prediction error (e.g. Akaike Information Criterion, AIC) is increasingly seen as a better approach for model selection. The main tool for model validation are residuals.

However, for state-space models, Pearson are not independent because of the dependence structure of the unobserved states. One-observationahead and one-step-ahead residuals are preferred. These can be formulated to be independent and Gaussian distributed. Key research recommendations involve when and how M can be estimated, how to include information about the precision of model inputs, and improved usage of diagnostics and model selection criteria.