statastan

statastan copied to clipboard

StataStan is the Stata interface to Stan.

StataStan comprises a command stan. If you are using Windows, you will also have to install the command windowsmonitor. Queries by email to Robert Grant at [email protected].

stan will fit a Stan model by Hamiltonian Monte Carlo. You can also ask for the posterior mode, which is found by optimization with the BFGS (or L-BFGS) algorithm. We intend to create more Stata commands:

• allowing CODA-style diagnostics and plotting after a model has been fitted and chains stored,
• fitting specific models given variables, matrices, globals -- in the style of the R package rstanarm,
• and a test script.

Current status update (from Robert)

Since version 16, Stata has featured close integration with Python, so you can type Python code inside your Stata script, and pass objects back and forth.

I wrote StataStan in 2014-15 and it is one of those make-do interfaces that are just wrappers for CmdStan. It relies on passing data, model, chains etc via text files, and that has always been a weakness, producing occasional problems dependent on users' OSs, security settings and so on.

My view now is that anyone who wants to use Stan from Stata should do so through PyStan or CmdStanPy. In interface terms, the speed and stability will be better than the -stan- command. I intend to freeze StataStan for Stata versions pre-16 but not to maintain compatibility with Stata going forward. It's hard to guess whether there will be need/demand to maintain compatibility with future versions of CmdStan, so post any such concerns here as issues, as they emerge.

Recent changes

Version 1.2.4, SSC release 4 March 2021:

• A warning is displayed if Stata version is 16.0 or later, advising the user to consider using PyStan or CmdStanPy via Python-Stata integration instead. This can be suppressed by the new 'nopywarn' option.
• System calls to copy and delete files are now done 'quietly'. Sometimes, the system reported "Error: file output.csv does not exist" and the like, which looked pretty alarming.

Version 1.2.3, SSC release 3 March 2017:

• Incorporates changes to Stan output variables from CmdStan (reads in columns after energy__)
• Bug in the load option fixed

Version 1.2.1 & 1.2.2, SSC release 16 Sep 2016, and hotfix 1 Oct 2016

• various bug fixes too tedious to list - see the commits for stan.ado for details

Version 1.2:

• You can run multiple chains with the chains option.
• The chainfile option can be abbreviated to chainf but no shorter. This avoids confusion with the new chains option.
• Whatever name you give in outputfile will now have ".csv" appended on the end. If you run more than 1 chain, you will get consecutively numbered files like output1.csv, output2.csv, etc. On Mac and Linux machines, the name should contain no spaces. We will try to work around this in future versions.
• To avoid clashes with existing files, both the working directory and the CmdStan directory get checked for pre-existing output*.csv files.
• The StataStan and CmdStan version numbers are displayed at the beginning of output. The CmdStan number will be used in future versions for back compatibility.

Version 1.1, SSC release 29 Feb 2016:

• print has been replaced with stansummary
• all files are cleaned up from the CmdStan directory and appear in the working directory instead
• added a 'keepfiles' option - without this, wmbatch (in Windows), winlogfile (in Windows), outputfile, .hpp will be deleted. The executable, datafile and chainsfile are retained no matter what.

Getting Started

1. Download and install CmdStan. Make sure you read the installation instructions and platform-specific appendix before installing. In particular, if you are using 32-bit Windows, you will need to add a file called 'local' to the 'make' folder before you run the make command, which should simply contain the text: BIT=32
2. Don't forget to specify the number of CPU cores when you build CmdStan, so you can take advantage of parallel chains and the faster model-fitting that can deliver.
3. Download the .ado and.sthlp files and save them in your Stata personal ado folder (click here if you don't know where this is)
4. Try out the different examples in the stan-example.do file, or under help stan
5. Try your own data and model. The Stan modelling manual is essential reading for this! Start small and build up. Options are listed in detail under help stan.
6. You can pass your current data (the stuff you see when you type browse in Stata) into Stan, but also you can send matrices and global macros, by specifying their names or typing 'all' in the matrices and globals options. Unlike BUGS / JAGS, Stan just ignores data that your model doesn't use.

Testing

StataStan has been tested with CmdStan 2.9 to 2.25, Stata versions from 11.2 to 16.1, and Stata flavors from IC to MP2. We have run it successfully on Linux (Debian), Mac (up to Mojave) and Windows (up to 10).

Other notes

• We find that specifying the CmdStan directory with a tilde in Mac OSX causes problems, and a complete path is advisable.
• On Mac and Linux, it is a really good idea to have the working directory and cmdstan directory paths without spaces. For parallel chains, this is essential.
• Non-existent globals and matrices, and non-numeric globals, get quietly ignored.
• Missing values are removed casewise by default (but you can change this)
• Users need to take care not to leave output file names as defaults if they have anything precious called output.csv or modes.csv etc. - these files will be overwritten.

Makers

StataStan was mostly made by @robertgrant with valuable contributions from @sergiocorreia, @felixleungsc, @syclik, @danielcfurr, Charles Opondo and @bob-carpenter, plus constant encouragement and strategic direction from @gelman. We also learnt a lot from John Thompson's Stata commands for BUGS. StataStan was initially partly funded (grant holders Andrew Gelman and Sophia Rabe-Hesketh) by the Institute of Education Sciences, USA.

Licensing

StataStan is licensed under BSD.