James Thorson1 and Melissa Haltuch1
January 9th, 2018 9:00 (PST): FSH 203
New features for spatio-temporal analysis using FishStats: An improved model for biomass sampling data, spatial expansion of age/length compositions, and short-term forecasts for distribution shift
In this talk, we quickly summarize the FishStats spatio-temporal toolbox (www.FishStats.org), including the spatio-temporal approach to index standardization. We then highlight three research topics since my last Think Tank in Nov. 2016. First, we highlight a new “Poisson-link delta model” (PLDM) that is more biological interpretable and parsimonious than the conventional delta-model that is widely used in fisheries science. We also discuss how this PLDM is interpreted as a computationally efficient approximation to a compound-Poisson-gamma model. Second, we discuss ongoing research using a vector autoregressive spatio-temporal (VAST) model to expand age and length composition samples, including a calculation for input sample size based on estimation variance. A simulation experiment shows that the VAST model has 20-30% lower errors than the design-based estimator, but a case-study application involving lingcod shows that small differences in composition expansion can result in a large difference in scale when used in an assessment model. Finally, we discuss ongoing research regarding short (1-3 year) forecasts of distribution shift using a climate-envelope or spatio-temporal model using retrospective skill testing. This shows that autoregressive spatio-temporal errors are important to generate unbiased forecasts with good forecast interval coverage. We therefore conclude by recommending increased use of the Poisson-link delta model, particularly for forecasting future distribution shifts, but further research regarding spatio-temporal expansion of composition data.