Ensemble models for data-poor assessment: the value of life-history information

Dr. Merrill Rudd1
April 03, 2018 9:00 (PST): FSH 105

Ensemble models for data-poor assessment: the value of life-history information

Scientists and resource managers need to understand a population’s biological parameters, such as mortality and individual growth rates, to successfully manage exploitation and other risks. In the absence of age data, growth rate estimates come from direct observation, proxies, or patterns in the length data while natural mortality rate estimates are often assumed based on empirical relationships with the growth rate estimates. Length composition of the catch can inform recruitment and fishing mortality, but parameter estimates depend on accurate growth and natural mortality rates. Uncertain or unreliable mortality and growth rates propagate high uncertainty in stock status estimates when relying on length data to inform vital stock assessment parameters. This study uses predictive stacking as a method of model averaging across a distribution of values for growth and mortality. We used the R package FishLife to develop distributions of life history parameters based on a multivariate model with taxonomic structure, drawing combinations of points from these distributions to integrate uncertainty in life history parameters into a data-limited, length-based stock assessment. Through simulation we demonstrate that predictive stacking leads to better assessment performance than assuming the parameter means from FishLife when the true values of life history parameters are unknown. We then applied the predictive stacking method for a U.S. Caribbean stock previously lacking accepted management advice due to debilitating uncertainty in life history parameters. This method will be applicable for stock assessments concerned with properly accounting for uncertainty in biological parameters, ranging from life-history-based to length-or age-based stock assessments.

Posted in Fisheries Think Tank.

Leave a Reply

Your email address will not be published.