Including parameter uncertainty in forward projections of computationally intensive statistical population dynamic models

The increased computational demands of modern statistical stock assessment models have made the standard methods to provide uncertainty estimates for forward projections impractical for timely results in many applications. However, forward projections and their associated estimates of uncertainty are an important and popular piece of management advice. We describe a less computationally intense method to estimate uncertainty in forward projections that includes both parameter uncertainty and future demographic stochastic uncertainty. This frequentist method uses penalized likelihood as an approximation to mixed effects and can be viewed as treating the future projection period as part of the estimation model rather than performing stochastic projections. This allows confidence intervals to be calculated using normal approximation based on the delta method. The method is tested using simulation analysis and compared with Bayesian analysis and with projections based on point estimates of the parameters. The method is applied to yellowfin tuna in the eastern Pacific Ocean.