Treating "nuisance" rate category parameters with low ESS

[dibasic]

Hi all,

I have GTR models for some data partitions, and some of the rate categories get low ESS values. This is probably due to the small number of nucleotides they contain making it hard to estimate the posterior for some types of substitutions.

I can simplify my model to HKY and this does something to clear them up, but what is the view on simply NOT estimating these parameters (ie unclicking the estimate flag in Beauti for the troublesome rate categories)? Is this a valid move?

thanks!

[Remco Bouckaert]

Hi John,

By not estimating some of the rate parameters in the GTR model, do you plan to use estimates from somewhere else? If so, you may end up with a bias towards the tree assumed in the method used to estimate these rates.

If you are not interested in the the rate parameters or other site model parameters, you may consider using the model test site model from the bModelTest package, which automatically reduces the complexity of the model if there is not sufficient data to justify richer models.

Cheers,

Remco

[dibasic]

thanks Remco. I have just installed bModeltest and will try that. If I have any questions about it I will get back to you.

[dibasic]

Hi Remco,

I used bModeltest and it works really well. Nice job!! I get high ESS for my likelihood, prior and posterior, and almost all other parameters. Questions: where can I find the site models that BEAST finally settled on for each of my partitions? Also, one of my BMT_rates parameters is still low — why is this still a problem and what can I (or should I) do about it?

cheers!!

[Remco Bouckaert]

Hi John,

The parameter with the low ESS you selected in the attachment is the sixth rate parameter. If most of the models do not use six parameters (that is, most of the time GTR is not the model being used) the sixth parameter is not being sampled, so you would expect a low ESS for that parameter and under these circumstances you do not have to worry about a low ESS. The ESSs for rates in the rate matrix (these have names starting with rateAC, rateAG, rateAT etc in tracer) should be good though.

Cheers,

Remco

[dibasic]

Awesome! I think I am done!! Many thanks for your help Remco. Bmodeltest is extremely handy.

[dibasic]

Can you just confirm where I need to look to see which sites model bmodeltest/beast ultimately chose?

[Remco Bouckaert]

o BMT_ModelInidicator is the index of the substitution model as listed in the appendix of the paper (http://biorxiv.org/content/early/2015/06/11/020792).

o substmodel is the model represented as a 6-digit number, where the position of the digit refers to rates ac, ag, at, cg, ct and gt respectively, and equal digits indicates that rates are shared, so 111111 is Jukes Cantor (if frequencies are kept equal), 121121 is HKY, 123456 is GTR etc.

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.