Adjusting for bias due to variability of estimated recruitments in fishery assessment models

Abstract

Integrated analysis models provide a tool to estimate fish abundance, recruitment, and fishing mortality from a wide variety of data. The flexibility of integrated analysis models allows them to be applied over extended time periods spanning historical decades with little information from which to estimate the annual signal of recruitment variability to modern periods in which more information about recruitment variability exists. Across this range of data availability, the estimation process must assure that the estimated log-normally distributed recruitments are mean unbiased to assure mean unbiased biomass estimates. Here we show how the estimation method implemented in the integrated analysis model, Stock Synthesis, achieves this unbiased characteristic in a penalized likelihood approach that is comparable to the results from Markov chain Monte Carlo. The total variability in recruitment is decomposed into variability among annual recruitment estimates based on information in the data and a residual variability. Because data are never perfectly informative, we show that estimated recruitment variability will always be less than the true variability among recruitments and that the method implemented here can be used to iteratively estimate the true variability among recruitments.