Stan Case Studies proposal

[Bob Carpenter]

I propose we start a new section of the web site titled "Case
Studies", which would host standalone documents meant to illustrate
particular models or data sets in Stan. I'm thinking something
like arXiv in its journal-like properties without the hassle.

We have one almost ready to go from Daniel Furr et al. at Berkeley
on the IRT 2PL model. I have two to offer, one on soil-carbon
modeling with ODEs, and one on biodiversity modeling using latent
discrete parameters (the Dorazio and Royle model). All three
reproduce widely used models in their fields.

All three of these were done in knitr, making them super-easy
to distribute in working form with HTML and LaTeX output.
I'd like to establish some standards like standalone data and
model files if we can, but this can perhaps evolve as we have
examples from more than two people.

Should we also include Jupyter or just plain old markdown
or just plain LaTeX or just plain HTML? Can we somehow
plug MathJax into markdown for the GitHub pages?

It'd be great if we could set up some kind of nice in-house
style for our HTML and LaTeX that would make me and Andrew and
Michael happy (I think we're by far the pickiest on these things
now that Malecki's left the circus, but all input welcome as
usual). The point is something that's both easy to read and
write, but also distinctive enough in look and feel that if
someone glances at it they say "hey, that's a Stan case study".

Each case study would get its own directory, and the index
pages would all be organized the same way a la arXiv.
We'd have at least the following fields,

TITLE: Soil-Carbon Respiration Models
AUTHOR: Bob Carpenter
DATE: October 2015
AREA: Biogeochemistry, ecology
TOPICS: ODEs with unknown initial conditions; latent mixtures
COMPARE TO: Sierra et al.'s R package soilR

CITE: Carpenter, Bob. 2015. Soil-Carbon Respiration Model.
<i>Stan Case Studies</i> 1(10).
http://mc-stan.org/case-studies/3

LINK: html formatted case study

REQUIRES: RStan 2.8

DOWNLOADS:
* case study itself (HTML, PDF)
* BibTeX citation (techreport)
* source code (.tar.gz, .zip of )
-- markdown
-- HTML
-- PDF
-- model files
-- R scripts (or Python or ...)
-- data files

The question is how much do we make this tutorial and how much
do we make it novel research? I don't think it matters --- my goal
is to have a nicely curated and indexed set of examples.

I think we can let people write these however they want. There'll
be overlap in intros to methodology, to Stan, etc., depending on
the level and audience they're written for, but I think that'll
be OK.

I thought I'd drop this here for feedback. Next step would be
to create a prototype and launch it on the web site, then ask for
user group feedback and contributions.

Another problem we're going to have is versioning of both Stan and
these documents. Each will be written against a current version of
PyStan or RStan or whatever, but these interfaces are likely to drift
over time in a way that'll break these demos. Backward compatibility
breaking has a steep cost for anything we want to maintain over time.

- Bob

[Michael Betancourt]

Why don’t we start with some actual examples to judge?

> --
> You received this message because you are subscribed to the Google Groups "stan development mailing list" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to stan-dev+u...@googlegroups.com.
> For more options, visit https://groups.google.com/d/optout.

[Bob Carpenter]

That's exactly what I'm proposing. I just wanted to see
if anyone had any suggestions or objections before I do it.

I can vouch for the content in all three of them. Mine
are here:

stan-dev/example-models/knitr

Daniel's are in our private ed repo.

- Bob

[Aki Vehtari]

Here's one methodological example
Projection predictive input variable selection using Stan+R
https://github.com/jtpi/rstan-varsel

Aki