Trouble with Skygrid while running beast

[idrissa dieng]

Hi dear All,

Hope this message will find you it peace and health?

I am running a Beast analysis on beast and now i am comparing model combinations;

All of them work perfectly but i stuck with UCLN + Bayesian Skygrid.

Parameters linked to Skygrid Log.posize didnt converge all ESS values are colored in red.

What should be the issues? and how to troubleshoot?

The oldest sequence i used is collected in 2012 and my sequences of interest are sampled in 2017 and 2018.

Best regards

[HS]

Hi,

I have the same problem with Bayesian Skyline model. Group sizes don't converge. In my case also ucld.std does not converge.

I tried a lot of things - changing parameter distributions, site and clock models, delta and weight of the group.size operator, etc. All of these does not help.

With a smaller dataset changing the Delta to 20 for the group.size operator worked. But with a bigger dataset it doesn't help. I remember many people wrote in this group about such a problem and there seems there is no a clear answer how so solve it.

I use Beast 1, but Beast 2 has gives this problem too. So dear Beast people, can you please have a thorough look in this issue?

Best,

Hovhannes

[Remco Bouckaert]

Hi Hovhannes,

There is no need for the group sizes or population sizes to converge. Usually, there are various ways to represent very similar population histories by different combinations of group and population sizes: consider a history with 2 epochs and constant population sizes in these 2 epochs. If you have 3 groups, then 1 group will have the same population size as one of the others, and one will differ. But there are 2 ways to do this, both of them resulting in the same population history. During MCMC, BEAST occasionally switches between these states, which then shows up as low ESSs for group and population sizes, but this is not something to worry about.

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 view this discussion on the web visit https://groups.google.com/d/msgid/beast-users/9b95fe42-d717-4dee-b4d3-acaad04c7c2fn%40googlegroups.com.

[Alexei Drummond]

Hi Hovhannes,

I would say that ideally you would get good mixing on all logged parameters. It may be okay not to in this case for the reasons Remco mentions, but I would at least run two independent runs and make sure that they give the same skyline plot. Of course there are various options to improve mixing including CoupledMCMC.

Cheers

Alexei

To view this discussion on the web visit https://groups.google.com/d/msgid/beast-users/84D9D032-6ACC-4703-AD09-C55D8AC787F3%40gmail.com.

[Remco Bouckaert]

PS you do want to make sure all other parameters have large enough ESSs and that you perform multiple runs so you can compare the demographic reconstructions from different runs and make sure they are similar.

On 9/09/2021, at 8:53 AM, Remco Bouckaert <higg...@gmail.com> wrote:

[HS]

Dear Remco and Alexei,

Thank you very much for important insights!

Do I understand correctly that if the number of groups will be chosen properly, then the problem of low mixing of the group.size parameters will evade? I tested many different numbers of groups and all of them resulted in low mixing too.

Shouldn't Bayesian Skygrid analysis help here?

Best,

Hovhannes