Bayesian estimation of non-Gaussian structural vector autoregressions (SVARs) via Stan. More specifically, this R-package aims at making Bayesian estimation of statistically identified non-Gaussian SVARs as easy as possible by providing an easy-to-use interface to cmdstanr along with various ready-made tools for the analysis of SVARs.
You must have cmdstanr installed (and properly set-up) on your computer before installing the package. However...
IMPORTANT: For now, cmdstan versions 2.32.1 and up do not work with the package, so you must use older release (but not older than 2.29.0)
Thus, in setting up cmdstanr you may follow the instructions in here, but with the exception of NOT installing the latest version of cmdstan. If you're getting confused and just came here for an easy-to-use R-package, don't worry, just follow the instructions below and soon all these unpleasant technicalities will be over and done with.
First, if you don't have cmdstanr R-package already installed, open a fresh R session and install cmdstanr by running the following command
install.packages("cmdstanr", repos = c("https://mc-stan.org/r-packages/", getOption("repos")))
Second, as required by cmdstanr, make sure you have a suitable C++ toolchain installed on your computer. What this means in practice?
On Windows: This means that you have to have Rtools installed on your computer. This is not an R-package and you may download Rtools installer for instance from here.
On Mac: This means that you must probably have at least Xcode command line tools installed on your computer. If you are not sure, the prompt in the next step should tell you if you need to do something.
On Linux: You probably already have what you need.
Third, make sure your toolchain is set up properly by running the following commands:
library(cmdstanr)
check_cmdstan_toolchain(fix = TRUE)
This should tell you if your toolchain is set up properly and it might also prompt you to accept any automatic fixes if this isn't the case. If you get stuck here, see this or this for more thorough instructions (or you can just carry on and hope for the best).
Fourth, install a suitable release (< 2.32.1) of cmdstan on your computer (not the cmdstanr R-package we have already installed!). For now it is recommended to use version 2.29.2 since this package has been most thoroughly tested with that version. You may install this version of cmdstan on your computer by running the following command (this might take a few minutes):
install_cmdstan(version = "2.29.2")
NOTE: If the above command fails when using Rstudio, try running the command without Rstudio. You can get back to Rstudio after succesfully installing a suitable version of cmdstan.
Now cmdstanr should be all set up. Even if you have an installation of the wrong cmdstan version (<2.29.0 or >2.32.0) on your computer (e.g. you have installed it earlier or by accident), bsvar() should find the right version and use that. However, if problems do persist, you may navigate to the location of cmdstan installations as given by cmdstan_path() and manually remove the folder(s) cmdstan-2.32.1, cmdstan-2.32.2... (or something similar) and see if that helps.
When you have cmdstanr all set up you are ready to install the package. Simply run the following commands:
if(!("devtools" %in% installed.packages())) install.packages("devtools")
devtools::install_github("jetroant/bsvar")
The installation of the package itself should be a breeze and only take a few seconds, however if there are a lot of dependencies to be installed (that is, if you do not already have other R-packages that are needed for this package to work installed on your computer) this might take a few minutes.
Everything should now be ready for you to use bsvar. First time you call bsvar(), a cmdstan program is compiled, which might take a few minutes. This however only needs to be done once, after which such speed bumps cease to exist. To test whether everything works as intended and to get that first time model compilation out of your way, you may run the following commands:
library(bsvar) # Loads the package
fit <- bsvar(y = matrix(rnorm(200), ncol = 2), lags = 2) # Estimates a bivariate model with two lags using random Gaussian data
prnt(fit) # Summarizes the parameter estimates
irfs(fit) # Plots impulse response functions
dist_plot(fit) # Depicts the estimated shock distribution parameters
These are by no means all the available tools included in the package, but hopefully give some idea of what can be done with the package. More elaborate examples will most probably be added here in the near future as the documentation of the package progresses.
Finally, a few words regarding the marginal likelihood estimation via the excellent bridgesampling R-package. bridgesampling::bridge_sampler can be called by runnning the following command:
marginal_likelihood(fit)
marginal_likelihood() is a wrapper for bridgesampling::bridge_sampler() that takes care of some pesky practicalities. Providing the stanfit object outputted by bsvar() directly to bridgesampling::bridge_sampler() will not work. This is mainly because (for now) bridgesampling only supports Stan models built with rstan, not with cmdstanr. To this end, marginal_likelihood() recompiles an rstan program based on the same code as the original cmdstanr program used for sampling from the posterior. For now, this needs to be once every session (annoying, I know!), so calling marginal_likelihood() for the first time in a session may take a few minutes as opposed to seconds later on. Also, very importantly, calling marginal_likelihood() requires an installation of rstan 2.26.0 or higher. You can check the version of your installation by running rstan::stan_version(). Check the instructions for updating your rstan installation here.