Model selection and data weighting methods for statistical catch-at-age analysis: Application to 1836 Treaty Water stock assessments

Recommended harvest limits for lake trout Salvelinus namaycush and lake whitefish Coregonus clupeaformis stocks in the 1836 Treaty Waters of the Great Lakes are based on statistical catch-at-age analysis (SCAA). The assessment models and methods are similar to those used to assess fish stocks in many of the word's major fisheries. My objective was to evaluate these methods with an eye towards suggesting improvements both for 1836 treaty waters and more generally. My results provide general guidance to stock assessment scientists with regard to data weighting and selecting among alternative assessment models. As a first step, I performed an analysis of the Lake Huron lake whitefish models' sensitivity to changes in "known" inputs and model structure, selected as examples of basic type of assessment used throughout treaty waters for lake whitefish and lake trout. All of the Lake Huron lake whitefish models were sensitive to changes in the methods used to estimate recruitment and time-varying selectivity, as well as to changes in their objective functions, and this indicated that further study of these aspects of the assessment methods was warranted.; Specifically with regard to the objective function, the assessment models were sensitive to changes in pre-specified variances associated with process and observation errors, which are used to weight the different data sources. This result is consistent with concerns expressed more broadly in the literature. I evaluated alternative approaches for estimating log catchability (process error) and log total catch (observation error) standard deviations within SCAA using Monte Carlo simulations: an ad hoc approach that tunes the model predicted log total catch standard deviation to match a prior value, and a Bayesian approach using either strongly or weakly informative priors for log catchability standard deviation. When process error variance is large relative to observation error (likely for many fisheries), reliable estimates of log catchability and log total catch standard deviations can be obtained in SCAA using a Bayesian approach with only a weakly informative prior on log catchability standard deviation.; The sensitivity of the Lake Huron whitefish models to the method used to model time-varying selectivity is also consistent with indications in the broader literature that SCAA assessments can be sensitive to misspecification of selectivity. I therefore evaluated four approaches for modeling time-varying selectivity within SCAA using Monte Carlo simulations: double logistic functions with one, two and all four of the function parameters varying over time, as well as age-specific selectivity parameters that all varied over time. None of these estimation methods out performed the others in all cases. In addition, I compared model selection methods to identify good (i.e., accurately matching the true fish population) estimation models. Degree of retrospectivity, the best selection method, was based on a retrospective analysis of bias in model parameter estimates as the data time series for estimation is sequentially shortened. I recommend this method of model section when considering different time-varying selectivity estimation approaches in SCAA.