next up previous contents
Next: Issues of Convergence Up: Sampling from the Posterior Previous: Metropolis Hastings

 

Gibbs Sampling

Efficiency of proposal density is an issue, but where the form of the full conditional distributions is known, these may be used to obtain proposals for the above algorithm.

The special case of the Metropolis-Hastings algorithm, where the proposal density, q is the product of full conditional distributions is called the Gibbs sampler. For example, consider the case of sampling from a target tex2html_wrap_inline2531 , with the knowledge of the conditional distributions, tex2html_wrap_inline2533 , and tex2html_wrap_inline2535 . Now, since tex2html_wrap_inline2531 = tex2html_wrap_inline2539 , detailed balance holds and the proposal is always accepted. In practice, it is possible that tex2html_wrap_inline2533 is known, but that tex2html_wrap_inline2535 has to be sampled using more general methods. In this case Gibbs sampling is combined with for example Metropolis-Hastings techniques. Such a sampling method is sometimes referred to as Metropolis-Hastings within Gibbs; although since Gibbs sampling is a special case of Metropolis-Hastings, this terminology is incorrect [9].



Cathal Walsh
Sat Jan 22 17:09:53 GMT 2000