A one-step-ahead pseudo-DIC for comparison of Bayesian state-space models

Area of Science:

• Ecology
• Statistics
• Computational Biology

Background:

• Conventional Deviance Information Criterion (DIC) assesses model prediction accuracy based on current state.
• Multivariate nonlinear Bayesian state-space models are crucial for ecological dynamics, like coho salmon populations.
• Existing DIC methods struggle with complex models and present computational challenges.

Purpose of the Study:

• Introduce a novel one-step-ahead DIC, termed DICp, for improved model selection in state-space modeling.
• Address the limitations of conventional DIC in distinguishing between competing nonlinear Bayesian state-space models.
• Provide a computationally feasible alternative for model evaluation in ecological studies.

Main Methods:

• Developed DICp, a modified DIC where prediction is conditional on the previous time step's state.
• Conducted simulations to compare DICp with conventional DIC for state-space model selection.
• Applied DICp to a multi-stage model of coho salmon abundance.

Main Results:

• Simulations demonstrated DICp's effectiveness in selecting appropriate state-space models.
• Conventional DIC produced misleading results, favoring incorrect models due to unaddressed process error inflation.
• DICp successfully eliminated the compensatory behavior of inflated process errors, offering a more accurate pseudo-DIC.

Conclusions:

• DICp provides a reliable and computationally efficient method for selecting between complex state-space models.
• The proposed DICp is particularly valuable for ecological applications where model mis-specification is common.
• This approach enhances the accuracy of population dynamics modeling and ecological forecasting.