Fisheries stock assessment models are used to estimate population demographics and dynamics such as abundance, biomass, and fishing mortality from input fishery data including total catch, composition of catch, and fishing effort. A goal of stock assessment is to accurately quantify population and fishery dynamics so stocks can be managed to achieve fishery objectives and long-term sustainability. Accurate and precise model estimates can be attained by using models and techniques that account for ecological complexity like variability in quantities across ages, years, or regions without overparameterization. The state-space framework is one such statistical technique that may allow for incorporating more stochasticity such that the model can better reflect reality. The state-space modeling framework assumes that unobserved "states" develop over time due to process error modeled as a random effect and that observed data have expected values based on these states but differ from expectations due to observation error.State-space stock assessment models (SSSAM) have experienced an outpouring of research and application in the past decade as computation processing power and novel software has facilitated the approximation of the high-level integrals necessary for SSSAM. SSSAM allows for several time-varying processes in recruitment, numbers at age, mortality, selectivity, and catchability, and has become an essential part of the contemporary fisheries modeling toolbox. With their swift advancement, it is important to understand best practices of applying state-space stock assessment models, and how data availability, the variability of process or observation error, and model structure may influence model results and accuracy.In Chapter 1 I built an age-based state-space stock assessment model that used fisheries dependent data, rather than fisheries independent surveys, as an index of abundance and was applied to Lake Michigan lake whitefish (Coregonus clupeaformis). The model predicted greater℗ abundance and lower mortality compared to the non-state-space model and domed rather than asymptotic selectivity. The state-space model also had reduced retrospective patterns in recruitment. Chapters 2 and 3 each used a simulation-estimation framework to generate catch and index data using a state-space stock assessment that assumed process variability in recruitment, expected survival (abundance), and selectivity. Simulations were based on a Gulf of Maine haddock (Melanogrammus aeglefinus) model and assumed different degrees of process and observation variance (Chapter 2) or assumed observation error likelihood distributions for the proportions at age (Chapter 3) to generate data. Simulated data were input into several estimation models with alternative assumptions about contributing sources of process variability and/or observation error distributions. The results show that state-space models which assume several sources of process variability can produce unbiased estimates even when processes are constant over time. The state-space models were able to estimate process variance in several stochastic processes under a broad range of true values. However, assuming variability in expected survival when it is deterministic can lead to the model not converging. Unbiased results are achieved when the observation likelihood is structured to account for inter-age correlation and overdispersion though such a framework may have difficulty allocating variance between process and observation sub-models.The concluding chapter places this work in the context of other age-based stock assessment models and argues for the inclusion of state-space in the modeling toolbox, as they can account for multiple time-varying processes and be used in a broad range of data contexts. This work provides a blueprint for where and how SSSAM may be best utilized in the future, particularly with data limited or data poor stocks and in cases where the process variance is unknown and should be estimated within the model.
Read
