11th National Stock Assessment Workshop (abstracts)
Specification of Observation Error Variances Grant G. Thompson; AFSC, Seattle, WA
Except for pure process error models, all stock assessment models require specification of observation error variances. However, there appears to be no consensus among practitioners as to how this should be done. One school of thought holds that the specified variances should be equal to the values implied by the respective sampling designs. A problem with this approach is that the distributional assumptions included in 'off the shelf' stock assessment packages may not correspond to the actual sampling designs. For example, most stock assessment packages assume that age/size composition data are drawn from a multinomial distribution, but actual sampling may violate the multinomial assumption. In such cases, it is necessary to compute a multinomial sample size that produces a variance equal to that from the actual sampling distribution. An example is provided. A second school of thought holds that the specified variances should be larger than those implied by the respective sampling designs, so as to compensate for any process error not included explicitly in the model. These larger values are typically determined within the stock assessment model itself by iterative reweighting. However, this practice is at best an approximation, as it can be shown that adjusting observation error variances cannot compensate completely for unmodeled process error. Moreover, this practice has the effect of adding parameters to the model, thus tending to increase the variances of estimates in general. It can be shown that better performance is obtained by modeling the process error explicitly.
Trawl Survey Designs for Reducing Uncertainty in Biomass Estimates for Patchily-Distributed Species Paul Spencer1, Dana Hanselman2, and Denise McKelvey1; 1AFSC, Seattle, WA; 2AFSC, Juneau, AK
"Patchiness" in the spatial distributions of marine populations such as Alaska rockfish can arise from heterogeneous habitat characteristics and can result in errors in survey biomass estimates when high-density patches are either over-represented or under-represented in survey trawls. In this study, we developed a spatial survey simulation model to evaluate the influence of spatial aggregation on biomass estimation and considered alternative trawl survey designs intended to reduce the variability of biomass estimates. Variants of double sampling procedures were simulated in which high-density areas identified from acoustic data in the first sampling phase were then assigned increased trawl sampling densities in the second sampling phase. Geostatistical analyses of hydroacoustic data collected in Alaskan trawl surveys were used to simulate spatial distributions of fish populations. Simulated survey biomass estimates and sampling variability were evaluated as functions of several factors, including the spatial aggregation of the population and sampling density. When the relationship between the hydroacoustic data and fish density was strong, the double sampling procedure resulted in reduced variance in estimated biomass relative to simple random sampling with equivalent sample size. However, the variance in estimated biomass from the double sampling design was not substantially reduced when the relationship between hydroacoustic data and fish density was weak. The potential improvement in variance when a strong relationship exists between hydroacoustic data and rockfish density offers motivation to continue to refine analyses of hydroacoustic data and rockfish spatial patterns.
