Evaluating methods for estimating the incidental capture of New Zealand sealions

Methods for estimating the incidental catch of sea lions in the SQU 6T squid fishery are compared. A simulate and test technique is used, with a trial model being used to generate synthetic data that has the same statistical properties as the actual capture data. The estimation approaches are then tried on the simulated data. This has the advantages that the simulated captures are known exactly, and that large numbers of simulated data sets may be produced.

The trial model is fitted to observer data from the 2000 to 2004 calendar years. The simulated data are derived by applying the trial model to all the tows from the catch effort database that were within SQU 6T and which targeted squid, from the same time period. The trial model was a generalised linear model (GLM) with fixed effects and with year and vessel within year random effects. The tested estimation procedures were a fixed effects GLM, a GLM with both fixed and random effects, and a simple ratio estimator. The ratio estimate assumes that, within any year, the sea lion strike rate is constant across the fishery.

In order to reduce sea lion capture, sea lion exclusion devices (or SLEDs) were introduced into the fishery in 2000. Since 2004 they have been used on most tows. The use of SLEDs has reduced the number of sea lions seen by observers, and consequently the estimation problem has become more difficult. The performance of the estimators is compared under assumed levels of SLED coverage, ranging from 10% to 90%.

An assumption of the models used is that the SLED retention probability has remained constant throughout the period of the data. This assumption may not be a good one, as re-fitting the trial model to data from 2002 to 2004 produces a model with a higher retention probability and a lower estimate of the total sea lion interactions in years when SLED use was high.

The conclusion of the analysis is that the model based estimators are indistinguishable and they both consistently outperform the ratio estimator.