Bayesian model selection for spatial capture-recapture models.

Area of Science:

• Ecology
• Statistical Modeling
• Computational Biology

Background:

• Ecological knowledge heavily relies on model selection for hypothesis testing.
• The increasing use of Bayesian hierarchical models in ecology renders traditional tools like AIC inapplicable.
• A lack of consensus exists regarding universal Bayesian model selection tools.

Purpose of the Study:

• To evaluate recommended Bayesian model selection tools for spatial capture-recapture (SCR) models.
• To assess the performance of these tools in selecting the 'true' model and estimating population size (N).
• To provide recommendations for model selection in Bayesian SCR analyses.

Main Methods:

• Applied and evaluated Bayes Factor (Gelfand-Dey, harmonic mean), Deviance Information Criterion (DIC), Watanabe-Akaike's Information Criterion (WAIC), and posterior predictive loss criterion.
• Utilized 25 variants of model selection tools.
• Generated 120 simulated datasets from a 'true' SCR model to assess model selection frequency and parameter estimation accuracy (N).

Main Results:

• No single model selection tool consistently achieved both accurate model selection and parameter estimation when data information content was low.
• The Bayes Factor (Gelfand-Dey method with MAP approximation) generally performed well when considering both objectives.
• Theoretically sound, albeit approximated, methods often outperformed newer, emerging tools like WAIC.

Conclusions:

• The Bayes Factor (Gelfand-Dey with MAP approximation) is recommended for Bayesian SCR model selection and parameter estimation.
• Model selection tool performance is contingent on data characteristics, particularly information content.
• Established, theoretically grounded methods can remain superior to newer approaches in applied statistics.