learning stab: what? how?

[Luis G]

Hi all,

I discovered Stan less than one year ago, almost at the same time as bayesian statistics, using Richard Mcelreath 'Statistical Rethinking'  and for last few months I have it to my work routine. In this thread, I would like to ask for some advice or someone who would share his or her experience on what should be the best way to learn it beyond a superficial level. Any tips or ideas would be welcome.

My main problem is that I am somehow trapped between easy problems, for which I don't need to use Stan, and difficult ones, that are too often too difficult- I guess this applies to most 'deeper level' languages. Most of my workflow are easy problems and for these, I use standard R packages, since these are much faster for exploratory analysis. Although I try to use stan as much as possible, it is just not practical given my time constraints. 

I am only driven to use Stan whenever I need more flexibility than it is provided by standard packages -typically hierarchical models that I need to customize for my purposes. Then the obstacle I face is a massive learning curve on many fronts: modeling, computational, etc,... : is my specification wrong? is my code inefficient? Am I trying to do something that just can't be done (maybe that's why there no ready to use package for it)? When I face these challenges, I feel I do not understand well enough what the software is doing for me. I typically plunge in the manual or forums.  These resources are great (the manual really is!), but massive. Too often, I end up giving up.

Does anyone have any suggestion on how to learn the language and what lies behind it in a more or less practical and systematic way? Is it worth, for instance, to spend some time reading about MCMC simulation techniques (some time in my bayesian course I learns Gibbs,Metropolis and had a sense of HMC)? Is it worth to read about data structures and language specific functions, for instance?

Anyway, I am just looking for you folks to share your experience with me.

[Andrew Gelman]

1.  I recommend rstanarm for the easy problems.  rstanarm has lots of functionality that regular regression software doesn't have, including draws from posterior, informative priors, regularization, and splines.

2.  To get into other uses, you can look at the example models and the case studies at the documentation on the Stan webpage.

3.  I recommend you focus on modeling and simulation of predictions, not on MCMC.  Stan does the MCMC for you.  At some point, to become a power-user it is helpful to understand Hamiltonian Monte Carlo, but this should not be your first priority.

A

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

[Michael Betancourt]

But if you _did_ want to learn more about MCMC and HMC

then check out https://arxiv.org/abs/1701.02434.

Also, I highly recommend that you check out the case studies

which are becoming really, really good,

http://mc-stan.org/documentation/case-studies.html.

[Luis G]

Thanks for the advice!