Beyond MCMC Black-Box: A Diagnostic-Driven workflow to Optimize Population Dynamics Models with NIMBLE

Abstract Bayesian hierarchical models are widely used in ecology to represent complex population dynamics and integrate multiple sources of information. However, their computational cost can limit their practical applicability, particularly for high-dimensional integrated population models. The NIMBLE framework offers flexible model specification in the BUGS language and several options to customize MCMC algorithms, that deliver significant gains in both the efficiency of posterior distribution exploration and computational time, without requiring the models to be recoded in another programming language. We develop a diagnostic-driven workflow to optimize MCMC sampling efficiency for Bayesian hierarchical models implemented within NIMBLE. The workflow combines node-level diagnostics (effective sample size, ESS per iteration and ESS per unit time) and systematic benchmarking of alternative sampler configurations to identify computational bottlenecks and guide the selection of more efficient samplers. We first demonstrate the approach using a series of toy models specifically designed to isolate key sources of MCMC inefficiency, including posterior dependence, parameter correlations, and interactions between posterior geometry and sampler design. We then demonstrate its applicability on an integrated population model used for international assessment of Atlantic salmon populations. Across both toy models and the realistic case study, targeted sampler optimization produces substantial efficiency gains, frequently exceeding one order of magnitude. Our results show that computational efficiency emerges from the interaction between model structure, posterior geometry, and sampler configuration, rather than from algorithmic choice alone. The proposed framework provides a reproducible and transferable strategy for optimizing MCMC algorithms in complex ecological models and other high-dimensional Bayesian systems.