Abstract:
Use of state-space modeling for stock assessments of salmon population dynamics has been limited compared to other sh and wildlife species, and this thesis aims to facilitate their application by investigating advantages of this approach, and resolving some methodological limitations. Other extensions such as the inclusion of abundance data from multiple stage-classes that have seldom been considered previously are also developed. A case study is explored where state-space, stage-structured models (`life-history models') are tted to coho salmon data, and advantages over traditional models are investigated. Unfortunately, selecting among a set of candidate models can be di cult for state-space models owing to the technical nature of some model selection tools, and choices about which part/s of the model should be focused on. It is shown that currently used measures of datalevel deviance could be improved by use of partially-marginalized deviance measures that allow di erences in process equations between candidate models to be detected. Whether the selected best model/s are adequate for inference is seldom assessed, or if it is, methodology is usually vague, and the power of detecting inadequacies has not been reliably assessed to date. It is shown that full assessment of model adequacy requires a partially marginalized extension of the more widely utilized posterior predictive checks (which are shown to have low power). Furthermore, the relative properties of test variables constructed based on di erent combinations of data and/or parameters/latent states are investigated. Finally, a signi cant limitation of state-space models in practice is the di culty in ensuring all parameters in the model, and most notably the variance components parameters, are identi able. Previous studies have attempted to constrain variance parameters by, for example, specifying the observation error variance as a constant after estimating sampling variation of the monitoring program that produced the abundance data. The latter technique was utilized herein, but before this was possible, methods for estimating the variance of spawner population size had to be developed. Novel estimators were developed to overcome this long-standing weakness of salmon monitoring programs, and a suite of estimators were compared in an extensive simulation study to determine the most robust methods to be applied when tting state-space models to salmon data.