The specification of the data model part in the SAM model matters

This paper considers a general state-space stock assessment modeling framework that integrates a population model for a fish stock and a data model. This way observed data are linked to unobserved quantities in the population model. Using this framework, we suggest two modifications to improve accuracy in results obtained from the stock assessment model SAM and similar models. The first suggestion is to interpret the "process error" in these models as stochastic variation in natural mortality, and therefore include it in the data model. The second suggestion is to consider the observed catch as unbiased estimates of the true catch and modify the observation error accordingly. We demonstrate the efficacy of these modifications using empirical data from 14 fish stocks. Our results indicate that the modifications lead to improved fits to data and prediction performance, as well as reduced prediction bias.