Tricks for working with JAGS

  • The Gamma distribution is frequently used as a prior for the precision in a model because of its conjugacy with the Normal. Given JAGS parameterization of the Gamma distribution with a rate and shape parameter, it’s not easy to select sensible hyper-parameters. It’s easier to think of hyper-parameters that correspond to the mean and the variance of a prior for the precision. To use mean and variance hyper-parameters (mu and sigma) with JAGS’s dgamma, use: dgamma( ((mu^2)/sigma), (mu/sigma)).
  • The glm module provides a series of more efficient samplers that JAGS automatically selects whenever possible (e.g. Albert/Chib (1993)). This usually leads to much better mixing. I tend to load the module right away with the actual rjags package, e.i. library(rjags); load.module("glm").
  • To check which samplers are selected by JAGS, use the list.samplers() function on the JAGS model object. This is useful, because sometimes a slight change in the coding helps JAGS to select better samplers. For example, when using a logit model in JAGS and you want the Holmes-Held sampler from the glm module to work, do not use dmnorm() but dnorm() for the prior of the coefficient, e.i. do not use blocking.
  • Avoid deterministic nodes as much as possible to increase speed. In particular, do all pre- and post processing in R including drawing from the predictive distribution and compute deterministic functions inside probabilistic statements, e.g. y ~ dnorm( (x*beta), 1) instead of mu .
  • For comprehensible / readable code, let the running index be the small letter corresponding to the capital letter of the upper limit, e.i. for(n in 1:N) instead of for(i in 1:N). This becomes especially useful when one has many nested hierarchies.

Gibbs Samplers in R and Notes

I uploaded several R files and notes for Gibbs samplers to my Website’s resource page. They include: Factor Analysis, Finite Mixture of Linear Regression Models and Probit Regression. I have written all this a while ago when I was reading about these models. Most of the samplers are fairly standard and well-implemented efficiently in R packages calling c++ in the background. These packages are also referenced in the R files. Nevertheless, the R code might be useful for pedagogical purposes or as building block for more complex schemes.