Convergence to extremely different areas of parameter space when used with Skygrid coalescent

[m-a-martin]

Trying to use CoupledMCMC on an analysis with a Skygrid coalescent model but the results seem to differ considerably from what would be expected.

As a simple example, I set up a BEAST2 analysis using the data in the tutorial here and ran with and without CoupledMCMC (2 chains, target swap probability of 0.05).

Without the CoupledMCMC the Ne trajectory is roughly comparable to that shown in the original tutorial.

However, with Coupled MCMC the population size parameters explore very extreme values that I would expect to be highly unlikely even given the heating (example of one of the Ne elements below).

This pushes the TreeHeight to really large values, on the order of 3E18.

This does not occur when using a constant coalescent with coupledMCMC, in which the estimated Ne is in line with expectations.

Possibly due to the log-transform on Ne in the Skygrid model?

xml.zip

[rbouckaert]

Hi Michael, thanks for the bug report -- The GMRFMultilocusSkyrideLikelihood in its current implementation is not compatible with CoupledMCMC. Notice that in the MCMC log there is an entry for SkyGrid and skyGrid.cutOff, but those are missing from the CoupledMCMC log, which is an indication the GMRFMultilocusSkyrideLikelihood is essentially ignored for the posterior when running CoupledMCMC. So, population sizes are sample from the prior and are not informed by the tree, which is what you observed.

[m-a-martin]

Hi Remco, thanks for the response and for confirming my observations. Appreciate the note on the log, I missed that. I'll table the use of coupled MCMC for this analysis for the meantime.