|Title||The utility of spatial model-based estimators of unobserved bycatch|
|Publication Type||Journal Article|
|Year of Publication||2019|
|Authors||Stock B.C, Ward E.J, Thorson J.T, Jannot J.E, Semmens B.X|
|Journal||Ices Journal of Marine Science|
|Type of Article||Article|
|Keywords||abundance; bias-variance tradeoff; bycatch estimation; discards; error; Fisheries; fishery; fishing effort; GAM (generalized additive model); Marine & Freshwater Biology; oceanography; random forest; rates; ratio estimator; spatial model; standardizing catch; strategies; US west coast groundfish fishery|
Quantifying effects of fishing on non-targeted (bycatch) species is an important management and conservation issue. Bycatch estimates are typically calculated using data collected by on-board observers, but observer programmes are costly and therefore often only cover a small percentage of the fishery. The challenge is then to estimate bycatch for the unobserved fishing activity. The status quo for most fisheries is to assume the ratio of bycatch to effort is constant and multiply this ratio by the effort in the unobserved activity (ratio estimator). We used a dataset with 100% observer coverage, 35440 hauls from the US west coast groundfish trawl fishery, to evaluate the ratio estimator against methods that utilize fine-scale spatial information: generalized additive models (GAMs) and random forests. Applied to 15 species representing a range of bycatch rates, including spatial locations improved model predictive ability, whereas including effort-associated covariates generally did not. Random forests performed best for all species (lower root mean square error), but were slightly biased (overpredicting total bycatch). Thus, the choice of bycatch estimation method involves a tradeoff between bias and precision, and which method is optimal may depend on the species bycatch rate and how the estimates are to be used.