Which assessment configurations perform best in the face of spatial heterogeneity in fishing mortality, growth and recruitment? A case study based on pink ling in Australia

Most fisheries stock assessment methods are based on the assumption that fish are homogeneously distributed across the area being assessed or that fish movement is such that local fishing pressure does not lead to heterogeneous spatial patterns of abundance. However, this assumption is seldom valid in practice. Seven alternative approaches for conducting assessments when confronted with possible spatial variation in fishing mortality, growth and recruitment are identified. These approaches range from ignoring spatial structure, to conducting a multi-area assessment that accounts for spatial variation in biological and fishery processes. These seven approaches are tested using simulations in which there is a single population with spatial heterogeneity, and the only linkage among areas is larval movement. The simulations are based on fishery and biological characteristics for pink ling, Genypterus blacodes, off southeast Australia. Non-spatial assessment configurations that aggregate spatially-structured data provide more precise, but nevertheless biased estimates of initial and final spawning biomass, as well as biased estimates of the ratio between initial and final spawning biomass. Assessment configurations that allow for spatial structure can provide imprecise and highly biased estimates, although these can be improved by changing the relative weighting applied to different data types. A spatially-structured assessment configuration that correctly matches the structure of the model used to generate the simulated data sets is unbiased but imprecise. When confronted with possible spatial heterogeneity in biological and fishery parameters, we propose conducting sensitivity analyses based on several model configurations to select the appropriate structure for an assessment. The capacity to examine model residuals spatially remains valuable for inferring problems with model specification.