adnuts

adnuts implements NUTS-based Bayesian inference for high-dimensional and complex hierarchical models by integrating with AD Model Builder (ADMB), Template Model Builder (TMB), and Stan.

Key Features:

• Integration with ADMB: Modifies the AD Model Builder (ADMB) source code to incorporate the no-U-turn sampler (NUTS) for Bayesian sampling.
• Integration with TMB and Stan: Links Template Model Builder (TMB) models directly into Stan to enable Bayesian inference via Stan's algorithms.
• Parallel sampling: Supports parallel sampling to improve computational efficiency for complex models.
• Laplace approximation validation: Allows testing the accuracy of the Laplace approximation used in TMB by comparing TMB approximations against NUTS-based Bayesian inference.
• Performance benchmarking: Provides reported case-study comparisons indicating ADMB and TMB performance generally within ±50% of Stan's speed.
• Flexibility across paradigms: Enables fitting models under both frequentist and Bayesian paradigms and testing the Laplace approximation for marginal likelihoods across arbitrary subsets of parameters.
• Extension to hierarchical models: Extends ADMB/TMB capabilities to perform Bayesian inference on hierarchical models that were previously infeasible with traditional frequentist methods.

Scientific Applications:

• Ecological hierarchical modeling: Enables Bayesian inference for complex hierarchical models commonly used in ecological research.
• Approximation and bias assessment: Facilitates validation of frequentist approximations and assessment of potential biases in parameter estimation by comparing Laplace approximations to NUTS-based inference.

Methodology:

Modifying ADMB source code to integrate NUTS sampling; directly linking TMB models into Stan for Bayesian inference; and utilizing parallel sampling for computational efficiency.