BEAST2: No ESS and very Low ESS for few parameters, Please Help!

[Ashok Mallik]

Hi all, 

I use BEAST2.  I tried running my analysis for divergence dating with 5 genes (74 taxa) with 5 fossil priors with following setting. 

I set all nuclear genes with GTR+G model and all mitochondrial genes with GTR+I+G (predicted by jmodeltest). 

I unlinked the models for each gene, linked the clock model and tree in "Partition" panel

Checked the estimate box for "Substitution rate" and "fix mean substitution rate"  in the "site model" and estimated "Clock rate" in clock model with relaxed clock log normal. All the parameters in Prior panel are kept as default with Birth Death model. 

After running 5 Million generations in BEAST I found that  ESS value are satisfactory  (few are low ESS), whereas running with 20 Million generations is very weird with NO ESS VALUE and very low ESS value in few parameters. 

Therefore, I request the users please suggest me what is happening in the result? Do I need to change few more settings to get proper result?

I have attached Screen Shot of tracer  both 5 and 20 Million generations here.

Thanks in Advance

Ashok

[Gustavo Sánchez]

Hello Ashok,

Please check the best model and which partition fit your data in the software "patitionfinder", then try to change your ucldStdev to uniform, all of them with 10 million mcmc and store every 1000.

If  you still have the same problem, please change the ucldStdev to lognormal. 

This recommendation if what I have got from the forum,

Hope it works

Best

[michaelm]

Hi Ashok,

some suggestions that may or may not solve your problem:

- Are you using BEAST v. 2.1.3 and BEAGLE? I've seen similar problems with this version, and I think I remember that using either BEAST v.2.2 or running it without BEAGLE solved the issue. But the developers surely know better whether or not this could be the reason.

- Even if jModelTest recommends the GTR+G+I model, many people in the field think this model is problematic as the estimates for the proportion of invariable sites and the alpha parameter of the gamma distribution influence each other and are difficult to estimate separately. Generally, GTR+G should work just as well if not better, and should lead to faster convergence.

- For the protein-coding mitochondrial markers, it may make more sense to split each marker according to codon position, and build three mitochondrial partitions: 1) all first codon positions combined, 2) all second codon positions combined, 3) all third codon positions combined. The reason is that substitution rates differ more strongly between codon positions than between markers, at least when compared within the mitochondrial genome.

- I guess I would always run a less parameter-rich model first just to test how fast and reliably BEAST converges. Using HKY instead of GTR would reduce the number of parameters of your analysis enormously.

Cheers,

Michael