inst/paper.md

title: 'The jagstargets R package: a reproducible workflow framework for Bayesian data analysis with JAGS' tags: - R - reproducibility - high-performance computing - pipeline - workflow - Make - Bayesian - JAGS date: "2020" output: pdf_document authors: - name: William Michael Landau orcid: 0000-0003-1878-3253 email: will.landau@gmail.com affiliation: 1 bibliography: paper.bib affiliations: - name: Eli Lilly and Company index: 1

Statement of need

Researchers who perform Bayesian statistics regularly experiment with models to refine, compare, and understand them [@bayesworkflow]. Each fitted model informs subsequent model-building choices, and even for experienced practitioners, the investigation often leads to final models that differ from the ones originally proposed. Fitting a model usually means applying Markov chain Monte Carlo (MCMC) or a similar method to approximate the full joint posterior distribution of the parameters [@bda3]. Flexible probabilistic programming languages such as JAGS have made model specification quick and straightforward [@jags], but computation time is still a bottleneck. A workflow can take several minutes or hours to run, and in subsequent iterations, researchers struggle to keep the results up to date with frequent changes to the models, code, and data.

Summary

The jagstargets R package [@jagstargets] is a workflow toolkit for Bayesian data analysis with JAGS. It helps users express a Bayesian statistical modeling exercise as a formal pipeline with dedicated steps for data generation, analysis, and summarization. Pipelines can be customized for popular use cases: for example, the analysis of a single dataset using multiple alternative models, or a large simulation study to test that a model is implemented correctly [@carpenter2017]. These pipelines, which can be visualized and executed using targets [@targets], support key features that increase the efficiency and reproducibility of Bayesian workflows. The steps of a pipeline are automatically orchestrated using optional distributed computing, and up-to-date tasks are automatically skipped if the upstream code and data did not change since the last run. Thus, researchers can quickly iterate on Bayesian workflows while maintaining agreement between the results and the underlying code, models, and datasets.

The jagstargets package is a bridge between packages R2jags [@R2jags] and targets. targets is a general-purpose pipeline toolkit for reproducible research and high-performance computing. It can be used on its own for Bayesian data analysis, but it does not natively support domain-specific functionality for Bayesian workflows or JAGS. jagstargets, on the other hand, leverages the existing capabilities of targets to more easily and concisely express computational pipelines for Bayesian data analysis, from single analyses or large-scale simulation studies. It abstracts away laborious error-prone tasks such as invoking JAGS, tidying posterior samples, and batching thousands of MCMC runs to improve computational efficiency, all of which reduce the user-side burden of writing custom R code for a targets pipeline. jagstargets is similar to the stantargets R package [@stantargets], the latter of which streamlines pipeline construction of Bayesian data analysis pipelines with Stan [@stan].

References



Try the jagstargets package in your browser

Any scripts or data that you put into this service are public.

jagstargets documentation built on Sept. 11, 2024, 6:21 p.m.