Practical Markov chain Monte Carlo.pdf

Practical Markov chain Monte Carlo Simon Jackman Stanford University /BASS February 3, 2012 Simon Jackman (Stanford) Practical Markov chain Monte Carlo February 3, 2012 1 / 1 BUGS and JAGS (Spiegelhalter, Thomas and Best 2000) and : Bayesian Inference Using Gibbs Sampling (Plummer 2010a): Just Another Gibbs Sampler, a look-alike to , with some subtle and welcome enhancements, runs natively on Mac/PC/Linux (hence my preferred software). I call from via the library (Plummer 2010b); I never ‘‘run’’ as a separate program. See §6.3. programs are not ‘‘procedural’’; they are ‘‘declarative’’, defining a model. Simon Jackman (Stanford) Practical Markov chain Monte Carlo February 3, 2012 2 / 1 / Workflow 1 develop model declaration syntax in a file (extension is ) 2 in , prepare data, pass data and model syntax to via the command 3 parses the data and model syntax, deduces the form of the implied DAG, creates a Gibbs sampler for the DAG 4 exploits conjugacy when it can recognize it; will use slice sampling, Metropolis etc when a conditional distribution is not recognizable in closed form etc. 5 or command actually runs the Gibbs sampler for a specified number of iterations 6 inspect results in ; ; Simon Jackman (Stanford) Practical Markov chain Monte Carlo February 3, 2012 3 / 1 JAGS Example Program for Regression model{ ## loop over the data for(i in 1:n){ ## form the conditional mean of y mu[i] - beta[1] + beta[2]*x[i] ## tau is inverse variance (precision) y[i] ~ dnorm(mu[i],tau) } ## prio