Performance of deviance information criterion model selection in statistical catch-at-age analysis

Most stock assessments involve fitting alternative models and selecting among them to provide management advice. Incorrect model specification can lead to unreliable population and mortality estimates, and methods to decide among assessment models so as to obtain reliable estimates are needed. We used Monte Carlo simulations to assess whether using deviance information criterion (DIC) model selection and averaging resulted in improved accuracy of important management quantities from statistical catch-at-age models. We challenged DIC with three estimation models (that differed in how they estimated catchability) and three scenarios of data accuracy and time-varying catchability. DIC usually selected the structurally appropriate model, and point estimates from the best model or the model average were relatively unbiased in that the average deviation from the true value was near zero. The distributions of point estimates about true values from DIC-based model averaging and from the best model (lowest DIC) were similar, perhaps because all of the estimation models were quite similar to the data-generating models. DIC seems to provide a useful metric to compare evidence in favor of alternative assessment models. This study is one of the first to evaluate the performance of DIC in models where the purpose is to predict unobserved quantities.