If you are looking for a post-doctoral position in statistical computing in 2012 then you may be able to help me with rewriting the coda package for R. But if you are interested, you don’t have much time to apply.The code base of coda is getting a bit old and needs refreshing. But coda has become an important piece of infrastructure for packages that use Markov Chain Monte Carlo methods. Even small changes to coda can have an impact on the many other packages that depend on it, so for some time I have avoided making major changes. The time has come for a ground-up rewrite. The new package will exist alongside the current coda package, which will be deprecated if and when we persuade everyone else to use the new version.
Here are the problems
- We need a new class for MCMC output.
- The mcmc class favours a time series representation of MCMC output. This is useful for convergence diagnostics, but not so useful if you want to do post-processing of MCMC output. Jouni Kerman’s rv package pioneered the idea of a class that could be used for transparent manipulation of objects representing random variables in R. I would like to incorporate these ideas into coda, while maintaining the time series aspect.
- We need a single class that can represent MCMC output, whether it comes from one chain, from multiple chains, or even from results pooled over chains.
- We need to update the statistical methods
- The methods in the coda package date back to the mid 1990s when the reviews written by Kate Cowles and Carlin (1995) and Brooks and Roberts (1998) were published. An update of the methodology is long overdue. The publication of the Handbook of Markov Chain Monte Carlo should be an excellent opportunity to bring the statistical methods up to date
- We need consolidation
- How many packages extend, or provide alternatives for, the methods in coda, and would the authors be happy to be part of a bigger consolidated package?
There is a fair amount of work here. It is more than I can do myself but I can supervise somebody else. This could be a project for an IARC post-doctoral fellowship in biostatistics. IARC fellowships are awarded on a competitive basis, and different groups at compete with each other. So there is no guarantee of success. If you are interested then contact me directly (plummerm at iarc dot fr) so that we can put in an application (applications are made jointly with the supervisor). The deadline for applications is 30 November.
I know there is not much chance of getting a response to this, but I am putting it out in case this matches somebody’s current need. If not, then there is always next year…