Modeling seabird bycatch in the U.S. Atlantic pelagic longline fishery: Fixed year effect versus random year effect

• We conduct a simulation study on how to modeling year effect for seabird bycatch.
• We compare performance of delta models with fixed and random year effects.
• Random-year-effect delta model performs better than fixed-year-effect delta model.
• Such a simulation is needed when annual estimates from models are inconsistent.

Year is usually modeled as a fixed effect in catch rate analyses because the annual variation is of interest. However, question rises when annual estimates are sensitive to whether modeling year as a random or a fixed effect. With the observer data from the National Marine Fisheries Service Pelagic Observer Program during 1997–2010, we conducted a simulation study using the delta model due to high percentage of zero observations in the observer data. The delta model consisted of two sub-models, one for modeling positive catch data, i.e., the longline sets with at least one seabird caught (positive catch sub-model) and the other for estimating the probability of catching seabirds (probability sub-model). We constructed five scenarios where data contained no year effect, fixed year effect, and random year effect with three increasing randomness, and evaluated the performance of three candidate models in terms of mean absolute error and mean bias. The three candidate models included the delta model where both sub-models had data select year based on its significance, the delta model where both sub-models fixed year in the model regardless of its significance, and the delta model where both sub-models modeled year as a random effect. Results showed that the model with random-year-effect performed the best in all scenarios for analyzing the positive catch data, followed by the one having data select year and the one with year fixed regardless of its significance. For estimating the probability of catching seabirds, performance of the three candidate models were competing in all scenarios except for one scenario where the probability sub-model having data select year performed the best. Combining the two sub-models, the random-year-effect delta model showed superiority over the other two candidate models for estimating seabird bycatch in the longline fishery. We suggest conducting such a simulation study in seabird bycatch assessment, especially in cases where yearly estimates from the random-year-effect and the fixed-year-effect models show great discrepancy.