This preview shows page 1. Sign up to view the full content.
Unformatted text preview: state.
6.2 Metropolis-Hastings algorithm The goal in MCMC is to construct a Markov Chain whose stationary dis
tribution is the posterior p(θ|y ). We now present the Metropolis-Hastings
algorithm. In addition to the distributions we have already used (likelihood
and prior), we will need a proposal distribution (or jumping distribution)
J (θ, θ' ) which will propose a new state θ' given the current state θ.
There are many options when choosing a proposal distribution which we will
discuss later. The proposal distribution will yield a random walk over the
parameter space, proposing steps θ → θ' . We accept or reject each step
depending on the relative posterior probabilities for θ and θ' . When we run
the random walk for long enough, the accepted values will simulate draws
from the posterior.
6.2.1 Some intuition into the Metropolis-Hastings algorithm Suppose we are considering the transition θ → θ' . If p(θ' |y ) is larger than
p(θ|y ), then for every accepted draw of θ, we should have at least as many
accepted draws of θ' and so we should...
View Full Document
- Spring '12