FISH 203
Monthly Archives: June 2024
Using Machine Learning with Fisheries Data
Using Machine Learning with Fisheries Data
Caitlin Allen Akselrud; Alaska Fisheries Science Center, NOAA
ABSTRACT TBD
When do environmental drivers improve advice from stock assessment models?
ABSTRACT TBD
Using simulation to evaluate alternative OA experiments for North Pacific crab stocks
FSH 203
ABSTRACT TBD
Dynamic structural equation models as novel framework for incorporating time-variation into stock assessment models
ABSTRACT: There is growing interest from fisheries agencies and managers to incorporate climate indices into stock assessment. Environmental covariates are typically included as independent variable(s) and used to predict variation over time in some stock-assessment parameter as a dependent variable (a “regression paradigm”). However, this practice has several drawbacks: (1) it ignores collinearity among potential covariates resulting from mechanisms linking different ecological processes; (2) it forces analysts to select between long-time series for physical drivers, and intermittent time-series for biological drivers (e.g., forage indices); and (3) it does not easily incorporate stakeholder or scientific knowledge about relationships among covariates.
In the first hour of this Think Tank, we will introduce dynamic structural equation models (DSEM) as alternative to this “regression paradigm” for including environmental indices as covariates. DSEM generalizes a wide range of linear and causal models, including dynamic factor analysis (DFA), vector autoregressive, integrated, moving average (VARIMA) models, and causal models. In particular, we emphasize that DSEM (1) allows analysts to incorporate scientific and stakeholder knowledge to model collinearity among covariates, (2) jointly imputes missing values based on covariate relationships, (3) can incorporate Bayesian priors to generalized qualitative network models, while also (4) allowing analysts to specify how ocean physics can affect stock productivity via intermediate mechanisms such as forage time-series. We will provide a short summary of the statistical background, and a hands-on code tutorial for R-package dsem.
Next, we will show how DSEM can be incorporated directly into state-space, age-structured stock assessment models to account for climate or ecosystem-driven variation in population processes. We present preliminary results for a case study on Gulf of Alaska pollock, where we are using process research and expert knowledge to develop candidate causal maps to describe the physical environment, foraging conditions, and ecosystem drivers governing annual recruitment success for this commercially important stock. We discuss the lessons we’ve learned so far while developing and refining alternative causal maps, and our most recent results from performance tests when compared to the operational stock assessment model.
View Recording Here: https://attendee.gotowebinar.com/recording/4692408206563899482