Accounting for uncertainty due to data processing in virtual population analysis using Bayesian multiple imputation
Virtual population analysis (VPA) is used in many stock assessment settings and requires a total catch-at-age data set where an age is assigned to each fish that has been caught. These data sets are typically constructed using ad hoc methods that rely on numerous assumptions. Although approaches are available to account for observation error in these data, no statistically rigorous methods have been developed to account for uncertainty from data processing. To address this, we investigated a Bayesian multiple imputation approach to filling missing size data. Using Atlantic yellowfin tuna (Thunnus albacares) and bigeye tuna (Thunnus obesus) as case studies, we evaluated the hypothesis that data processing is as important in determining management reference points in stock assessments as conventional sources of uncertainty. Size imputation models accounting for location, season, and year provided good predictive capacity. Uncertainty from data processing could be large; however, the circumstances for this were unpredictable and varied depending on the stock. These results indicate that VPA assessments should attempt to account for uncertainty in data processing to avoid potentially large compression of uncertainty in assessment results.