Abstract
Abstract
Fishery-dependent indices of abundance are often used to inform stock assessments, but fishing effort is generally not randomly distributed. Fishers concentrate sampling in locations expected to be profitable, leading to a form of preferential sampling that can bias model-based indices of abundance. We show that when the strength of preferential sampling is constant over time, it biases only the magnitude of an index, which is absorbed by estimated catchability parameters in most stock assessments. A more interesting case arises when the strength of preferential sampling changes over time; in these cases the index trend can be biased in a way that conventional catchability estimation cannot correct. Using simulations structured on long running trawl survey data from the west coast of the USA for shortspine thornyhead (
Sebastolobus alascanus
), we evaluate two structurally different methods for correcting time-varying trend bias: a joint spatiotemporal model of the density and sampling processes, and a discrete-choice model of fisher location choice. Our simulations include two mechanisms generating preferential biases: effort tracking local density, and effort responding to a spatially structured, but unobserved anomaly that also inflates catch. We show that the ability of the estimation methods to correct bias depends on the mechanism responsible for concentrating effort; while the joint model is robust to both data-generating mechanisms, the discrete choice model is only able to correct biases when the mechanism is anomaly-driven. As a real-world case study, we apply these corrections to simulated preferentially sampled data and include indices in the shortspine thornyhead stock assessment model. We found that uncorrected time-varying preferential sampling positively biased estimates of a major recruitment event and, to a lesser extent, unfished recruitment. This bias propagated through to bias in overfishing limits that grew as the population was projected forward. While bias-corrected indices still produced some positive biases in overfishing limits, these were substantially reduced compared to the naive approach. As the mechanism of preferential sampling is rarely known in advance, our results suggest that the joint model represents a robust default for correcting fishery-dependent indices of abundance.