Writing specifications for papers using Stan for simulations

  • I am using the Stan program for MCMC simulation.I want to know some of the writing conventions for paper writing after using the Stan program.At the same time ,I also want to know that how to write mcmc process in the paper to publish.
    Thank you very much.
1 Like

Hi,

look at some previously published papers in your particular field to see how people have done it before. If there aren’t any, then I’d look at other papers to get some inspiration. My current favorite is a paper by @richard_mcelreath et al.

THE LIFE HISTORY OF HUMAN FORAGING: CROSS-CULTURAL AND INDIVIDUAL VARIATION

It wouldn’t surprise me, however, if @betanalpha has some recommendations somewhere on how to report Bayesian analysis in research publications (if not then I would hope he has it in his pipeline :)

3 Likes

John Kruschke has written about what to report for a Bayesian analysis in general (not Stan specific) in both editions of his book Doing Bayesian Data Analysis In the first edition, see pages 508-510. In the second edition, see pages 721-724.

Also see http://doingbayesiandataanalysis.blogspot.com/2012/05/how-to-report-bayesian-analysis.html

2 Likes

To be totally honest all of my applied work is done with collaborators who manage the politics of how to make a Bayesian analysis palatable to their field. Exactly what information editors and reviewers expect varies strongly from field to field and often is based more on field-specific conventions than anything statistical.

From a statistical perspective I would personally want to see a link to the data, the exact configuration of Stan used to call the sampler (including random seed), the version of Stan, the interface and version of that interface, the operating system, and then all diagnostics that were considered and what thresholds were used. In other words everything that would be needed to recreate the analysis if needed.

5 Likes

If you want bit-level reproducibility, you also need the same hardware, same OS version, and the same compiler, version, and settings.

For a more stochastic notion of reproducibility, it’s enough to have the data and model and a version of Stan that runs the model.

2 Likes

Docker has been solving this for me. There I can pick exactly which OS, compiler, etc. I want.

1 Like

I’ve given up on bitwise reproducibility – I can’t even guarantee reproducibility when going from an RStudio terminal to Rmarkdown parsing all I want is to be able to identify what version of the codes were used so I can at least identify any incompatabilities that might arise.