next up previous contents
Next: Graphical Representation of a Up: Statistical Methodology Previous: Issues of Convergence

 

Nature of Multidimensional Posterior

Another factor which can seriously affect the rate of convergence, and the mixing of the chain is the shape of the posterior distribution. Jarner [20] has considered in detail the theoretical speed of convergence for different shapes of the posterior, his work building on theoretical work of Roberts [43]. This phenomenon was observed in practice, in the context of the sampling in the models presented later on.

An example demonstrates what occurred. Consider a two dimensional posterior, tex2html_wrap_inline2583 which is a ridge, as illustrated in Figure gif. The methods described sample the posterior, by starting at some location in the plane, and moving first in the x direction, then in the y direction. Each move is either accepted or rejected. It is clear from Figure gif that the methods will take a substantial amount of time to move from the lower left to the upper right of the target density. In Section gif it is mentioned that the chain took a very large number of steps to examine the posterior, and that this difficulty could have been avoided. The difficulty presented is that the chain takes many steps to sample from the complete support of the distribution; that is the chain does not mix well.

  figure453
Figure: Posterior Density in x and y.   

  figure461
Figure: Transformed Density with Better Mixing in tex2html_wrap_inline2589 and tex2html_wrap_inline2591 .  

A reparameterisation of the problem allows a much quicker sampling strategy. Consider the reparameterisation shown in Figure gif. The transformation tex2html_wrap_inline2593 yields a basis for which the density is now a ridge running in the direction of tex2html_wrap_inline2591 only, with width along the tex2html_wrap_inline2589 direction. The samplers now have the ability to traverse the target in just one step of z.

Of course, in order to carry out such a reparameterisation, one needs to have knowledge of the posterior distribution. As such, the transformation may not in general be specified a priori. It therefore requires an initial running of the chain with untransformed variables and based on this a transformation can be made. In the case of the models discussed later it was found that the transformation speeded up convergence by an order of magnitude.


next up previous contents
Next: Graphical Representation of a Up: Statistical Methodology Previous: Issues of Convergence

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