Abstract:
People often need to customize statistical models for particular problems and then consider a variety of methods for estimation and inference. Customizations may include adding components across space, time, repeated sampling, networks, non-parametric relationships or distributions, or multiple data sources, among others. Methods may include MCMC with potentially many kinds of samplers, empirical Bayes or marginal maximum likelihood, Laplace approximation and its extension to adaptive Gauss-Hermite quadrature, integrated nested Laplace approximation and related methods, sequential Monte Carlo, and others. Some methods represent hybrids, such as Particle MCMC combining particle filtering and MCMC. I will give an overview of the NIMBLE framework (R package nimble) for such problems. NIMBLE combines a language for writing models (an extension of the BUGS/JAGS language) and an algorithm programming system from R, in which all built-in algorithms are written and users can write new algorithms. Models and algorithms are automatically generated into C++ and compiled, and they can use derivatives of arbitrary order. NIMBLE has been used in over 600 peer-reviewed publications across many fields, with a center of gravity in ecology and environmental science. I will illustrate NIMBLE with a 2-parameter logistic item-response theory model with nonparametric distribution of abilities for education and health data and with a multi-species occupancy model of bird species distributions in California. The first example shows use of Bayesian nonparametric distributions, while the second shows configuration of MCMC samplers such as Barker, slice, and HMC along with comparisons to Laplace and INLA-like nested approximations. Finally I will point to future developments for NIMBLE, including greater scalability, better workflows, and better use from other packages.
Perry de Valpine is a mathematical and statistical ecologist. His research interests include population dynamics, theoretical ecology, and computational methods for fitting biologically realistic models to data. He is a lead developer of NIMBLE (r-nimble.org), a flexible computational system for hierarchical statistical modeling. He has contributed to modeling and data analysis of many systems in environmental science, including bird communities, large carnivore populations, agricultural insect dynamics, forest change, soil microbiomes, plant chemical diversity, fisheries, and more. De Valpine is a Professor in the Department of Environmental Science, Policy, and Management and teaches courses on statistical and mathematical modeling methods in ecology and environmental science.
