sky grid.cutoff ESS=1

[Luca]

Dear BEAST users,

I'm trying to use the BEAST on my dataset, about 240 viral sequences of 400 nt length. I applied the GTR+G+I substitution model and both constant and relaxed clock type. Using the Constant size, exponential growth, GMRF and Skyride. I run the BEAST for 50 million iterations, for the constant size and GMRF I didn't see an Increasing of ESS in the different parameters. I have low ESS values (20/25) but better than the other models for the SKyride putting as cut off 4500 and 50 as grid-points because one article only previously described the age of the root at 4500 years ago. But if I use 20 grid-points and 2000 as cut-off I see a better increasing of ESS values (56). As you can see from the image all the ESS increase except the sky grid.cutoff ESS that is always stopped on 1. How can I choose better the sky grid cutoff value?

In addition, If I run the BEAST for 100 or 150 million iterations I think I will reach the 200 ESS values in all parameters, the problem is that at 50 million the ESS values for ac ag at gt ct cg are around 3000 so if run for 100 million or more I will reach more than 10000 ESS values. What can I do?

Thank you.

all the best, Luca.

[Philippe Lemey]

Hi Luca,

Low ESSs for pop size parameters, rootHeight and clock rate frequently indicate weak temporal signal. This may also explain different mixing behavior under different coalescent models. The alignment length seems rather small, so assuming you have dated tips, it may be useful to investigate the temporal signal using TempEst. More informative priors (e.g. on the rate or rootHeight) should help if the sequence data cannot inform the dated tip model very well.

As the skygrid cut of value is not a parameter that is estimated, but something that needs to be fixed a priori, it will always have an ESS of 1, so that is not a problem. Concerning the value that one should choose for this cut-off, a preliminary run using a simple coalescent model should give a reasonable indication of the rootheight. According to your Tracer summary, something around 1500 could be a reasonable choice. There is no problem with very high ESSs for your the substitution parameters, but you could reduce the weight on their operators, or increase the weight of the parameters for which mixing is far more challenging (pop size parameters, rootHeight and clock rate).

Best,

Philippe

--
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.
<Screen Shot 2018-06-12 at 16.26.17.png>