Abstract
State-space models have been promoted as the next-generation of fisheries stock assessment and evaluation of their reliability is needed. We simulated operating models that varied fishing pressure, magnitude of observation error, and sources of process error. For each operating model, we fit a range of estimating models with correct and incorrect configurations. We measured reliability of estimating models by convergence rate, accuracy of AIC-based model selection, estimation bias, and magnitude of retrospective patterns. All reliability measures were generally better with lower observation error, contrast in fishing pressure over time, and when median natural mortality rate is known. The magnitude of the log-likelihood gradients was not a reliable indicator of convergence. AIC can generally distinguish process error source with lower observation error and higher true process error variability. Distinguishing the stock recruit relationship with AIC required large contrast in spawning biomass and low recruitment variation, but bias in stock-recruit parameter estimation was prevalent. Retrospective patterns were not large for mis-specified models. These findings improve our understanding of when results from state space models will be reliable.