[MUSIC] So far we've demonstrated MCMC for just one single parameter. What happens if we seek the posterior distribution of multiple parameters and that posterior distribution doesn't have a standard form. One option is to perform Metropolis Hastings by sampling candidates for all the parameters at once. And accepting or rejecting all of those candidates together. While this is possible, it can get complicated. Another simpler option is, to sample the parameters one at a time. As a simple example, suppose we have a joint postilion distribution for two parameters, theta and phi, given our data Y. And let's suppose we only know this up to proportionality. We are missing the normalizing constant. What we've calculated is the function g of theta and phi. If we knew the value of phi, then we could just draw a candidate for theta and use this g function to compute our Metropolis Hastings ratio and possibly accept the candidate. Before moving on to the next iteration, if we don't know the value of phi, then we could perform a similar update for it. We would draw a candidate for phi using some proposal distribution. And again, use this g function where we plug in the value of theta. To compute our Metropolis-Hastings ratio. We pretend we know the value of theta by substituting in its current value or current iteration from the Markov chain. Once we've drawn for both theta and phi, that completes one iteration and we begin the next iteration by drawing a new theta. In other words we are just going back and forth,updating the parameters one at a time,plugging in the current value of the other parameter into the g function. This idea of one at a time updates is used in what we call Gibbs sampling. It also produces a stationary Markov chain, whose stationary distribution is the target or posterior distribution. If you recall, this is the namesake of JAGS, which is Just Another Gibbs Sampler. Before describing the full Gibbs sampling algorithm there's one more thing we can do. Again using the chain rule of probability. We know that the joint posterior distribution of theta and phi can be factored. First into the marginal. Posterior distribution of phi times the full conditional distribution of theta given phi and the data. Notice that the only difference between this full joint posterior distribution and this full conditional distribution here, is multiplication by a factor that does not involve theta at all. Since this g function when viewed as a function of theta is proportional to both the full posterior and this full conditional for theta. We might as well have replaced g with this distribution when we performed the update for theta. This distribution of theta given everything else is called the full conditional distribution for theta. Why would we use it instead of g? In some cases the full conditional distribution is a standard distribution that we know how to sample. If that happens, we no longer need to draw a candidate and decide whether to accept it. In fact, if we treat the full conditional distribution as a candidate proposal distribution, the resulting Metropolis-Hastings acceptance probability becomes exactly one. Gibbs Samplers require a little bit more work up front because you need to find the full conditional distribution for each parameter. The good news is, that all full conditional distributions have the same starting point. The full posterior distribution. So, using the example above, we have that the full conditional distribution for theta, given phi, and y, will be proportional to full joint posterior distribution of theta and phi, given y. Which is also proportional to this g function up above. Here we would simply treat phi as a known constant number likewise the other full conditional will be phi given theta and y. Which again will be proportional to the full joint posterior distribution, or this g function here. We always start with the full posterior distribution, thus the process of finding full conditional distributions, is the same as finding the posterior distribution of each parameter. And pretending that all of the other parameters are known constants. The idea of Gibbs sampling is that we can update multiple parameters by sampling just one parameter at a time and cycling through all parameters and then repeating. To perform the update for one particular parameter we substitute in the current values of all the other parameters. So let's call this our Gibbs sampling algorithm. Here's the algorithm. Let's suppose we have the joint posterior distribution for two parameters, phi and theta, like we do here. If we can find the distribution for each of the parameters given all the other parameters and data, the full conditional distributions, then we'll take turns sampling the distributions. The first step in the Gibbs sampler will be just like the first step in Metropolis Hastings where we initialize. So we'll start with a draw for Theta not and phi not. The next step is to iterate so for I in 1 up to M we are going to repeat the following. The first thing we'll do Is using the previous iterations draw for Phi. So Phi, i -1, we're going to draw, Theta i from it's full conditional. By plugging in the old value of phi. Then, once we've completed this draw for phi. I'm sorry, this draw for theta i. We're going to use it. So using Theta i, the most recent draw for theta, we're going to complete a draw for phi i using its full conditional distribution. And we're going to condition on theta i. Together, these two steps complete one cycle of the Gibbs sampler and they produce a pair. We'll get a theta i, phi i pair. That completes one iteration of the MCMC sampler. If there are more than two parameters we can handle that also. One Gibbs cycle would include an update for each of the parameters. In the following segments we're going to provide a concrete example of finding full conditional distributions and constructing a Gibbs sampler [MUSIC]