MCMC with multiple chains and starting points

[aliakhavan89]

I couldn't find the explicit definition in the tutorial, so I'm posting my question here: When running a standalone MCMC with multiple chains per each parameter, do all chains start from the same starting point?

[tomfid]

No. There are three initialization methods: overdisperse, underdisperse, and hybrid (set by the MCINIT parameter).
- over draws in the same way Powell multistart works, from a very broad powerlaw distribution (or uniform if both bounds are set); in the future, I'd like to modify this to draw from your prior (if any)
- under starts from the base simulation parameters, plus a small epsilon
- hybrid starts from the base parameters, plus an attempted move in a random direction

[aliakhavan89]

Thank you, Tom! I think I got confused thinking all chains start from the same point and MCINIT controls the level of "added noise" for sampling. But noise is clearly controlled by other parameters, so not sure why I got confused. Thanks a lot for the clarification!

[tomfid]

One way to assess what's going on is to look at the _distrib file that gets generated. If you filter out the final sample and plot a parameter against iteration, you can see how the distribution evolves.