multiple runs-more iterations or different prior parameters?

[quattrinia]

hi there,

i am having difficulty assessing issues with convergence from a divergence dating analysis in beast.

i am running 8 independent MCMC runs on the same dataset for 15Mil generations each, and then combining these in log combiner. my ESS values are really low for most of my parameters, and combining the files in Log Combiner yields even worse results. i know that I need more iterations, but before I do this,  I am also wondering if there is an issue with my priors -and how to go about assessing this? any help would be appreciated!

My dataset includes ~235 taxa and 25 loci. I have a starting tree with estimated node dates in my .xml file, and priors include 8 fossil calibrations, a relaxed lognormal clock, and a yule tree. 

i have attached the output from one run with associated xml file, and a log file from 5 runs combined. 

thanks in advance,

andrea

[Remco Bouckaert]

Hi Andrea,

It looks like you need to run the analysis much longer to get convergence. One thing  that may speed things up a little is putting an upper bound of 1 on the ucldStdv, which looks like will end under 1 eventually, but spends most of the chain much above that. 

Since you are combining MRCA priors with the Yule model, expect some interaction between these tree priors. To asses how this happens you can take a sample from the prior (tick “Sample from prior” box in the MCMC panel in BEAUti) to generate an XML file, then run BEAST on that file and inspect the results in Tracer.

Cheers,

Remco

--
You received this message because you are subscribed to the Google Groups "beast-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to beast-users...@googlegroups.com.
To post to this group, send email to beast...@googlegroups.com.
Visit this group at https://groups.google.com/group/beast-users.
For more options, visit https://groups.google.com/d/optout.
<andrea.zip>

[quattrinia]

Ok. thanks~ I am glad it was as simple as that :)

Best,

Andrea

[quattrinia]

hi remco,

I have tried running this again with more generations (90mil), and ESS values are not really changing at all for some parameters (ucld mean, ucld stdev, birth rate, etc., although ESS values are > 200 for most of my MRCA, rate, and freqparameters. I have also sampled from the prior as you suggested. Can this output from "sampling the prior"  (og1.prior.log) help me troubleshoot which parameters I need to change-and how I should change them? 

Thanks in advance~

Andrea