When Sampling from the Prior, should it converge?

[Dylan O'Hearn]

I'm using the "sample from prior" option in Starbeast3 with a dating analysis to ensure that my fossil calibrations are adequately represented.  The node ages converged well and match what I intended, however, I noticed that the "posterior" of the prior just wanders around without ever setting on any particular value, and this seems to be driven mostly by the speciescoalescent and vectorprior, which both show the same pattern.  Does this indicate a problem with my priors, or do only my specific parameters of interest matter? 

[Jordan Douglas]

Hi Dylan,

This poor mixing does not indicate there is anything wrong with your priors.

Given an enough time, the MCMC will eventually converge. However the prior often takes significantly longer to converge than the posterior for a model this complex. The multispecies coalescent contains many different trees, which are correlated with each other through tree prior parameters (eg. population size) and of course the species tree. The full space of trees is very large and very difficult to traverse, especially in the absence of data to "guide" the MCMC. 

If the node ages are the only thing you care about here, then having a good ESS for them should be fine. But to be safe you may want to keep the chain running until the prior ESS also increases just in case there is another undiscovered optimum.

Jordan