HyperPrior parameters don't converge

[philipp r.]

Hi BEAST users,

I have troubles with the convergence of the Hyperprior parameters of the tree model in my BEAST run (see attached pictures).

My run is an 8 gene concatenated analysis, with substitution models estimated in JModeltest.

The clocks are unlinked with one gene having a strict clock. The root has a fixed age.

I use a Birth Death tree model, however I had the same problems when using Yule.

Increasing runtime doesn't seem to help, I logcombined several 500 mio runs without success.

The other parameters converge well though, after around 50 Mio generations.

I would be very happy for any input on this, I am running out of idea what I can do to improve convergence.

Thank you,

Philipp

[Alexei Drummond]

Dear Philipp,

Did you really mean to use a hierarchical model here? Why are you estimating gamma hyper prior parameters on the birthRate and relativeDeathRate parameters? Shouldn’t you be using two fixed gamma (or lognormal) priors on the birthRate and relativeDeathRate parameters? 

If you really want to estimate the gamma hyper parameters then you should definitely not have uniform priors on the alpha and beta parameters of the gamma hyper prior. You can see from your output that the gamma hyper prior alpha and beta parameters are all very high (> 1e11). Did you really want to explore that part of parameter space?

Perhaps you have introduced the gamma hyper priors in BEAUti by mistake?

Cheers

Alexei

--
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 http://groups.google.com/group/beast-users.
For more options, visit https://groups.google.com/d/optout.
<screen1.png><screen2.png>

[philipp resl]

Thank you Alexei for your answer. Yes it was a mistake, in fact I played around quite a lot with different parameters and models, and reused the same XML file again and again. At some point I have probably overlooked the extra parameters. I discovered this already shortly after I sent the message to the user group and created a fresh XML file. Now the analysis seems to converge well.

all the best and thank you again for your help,

Philipp