One size does not fit all: Customizing MCMC methods for hierarchical models using NIMBLE.

Area of Science:

• Ecology
• Computational Statistics
• Bayesian Inference

Background:

• Markov chain Monte Carlo (MCMC) is crucial for Bayesian hierarchical models.
• Increasing model complexity necessitates improved MCMC efficiency.
• Ecological models often require computationally intensive statistical analyses.

Purpose of the Study:

• To evaluate strategies for enhancing MCMC efficiency in ecological models using NIMBLE.
• To compare NIMBLE's performance against JAGS for various model formulations and sampling strategies.
• To determine how MCMC efficiency is influenced by model structure, size, data, and sampling methods.

Main Methods:

• Utilized NIMBLE (R package nimble) and JAGS for MCMC analysis.
• Applied common ecological models: species occurrence, abundance, multiseason/multispecies occupancy, and N-mixture models.
• Compared sampling discrete latent states versus integrating over them, and univariate versus block-sampling.

Main Results:

• For simple models, computational approaches showed minimal differences.
• Model complexity revealed strong interactions between formulation and sampling strategy impacting MCMC efficiency.
• NIMBLE consistently outperformed JAGS, especially for complex models (10-12x more efficient).

Conclusions:

• No single MCMC strategy fits all; problem-specific approaches are essential.
• NIMBLE is a valuable tool for ecologists, particularly for complex Bayesian models where JAGS is slow.
• Further guidelines and customizable approaches are needed for fitting hierarchical models effectively.