Help with ESS on calibration (ucld Stedv and ucld mean)

[Gabriela Cruz Luna]

Hi everyone

I have a very quick question, nope someone could help.

I'm trying to do a calibration using fossils. 

My matrix has 329 taxa and 1039 bp.

I have run my job twice as following:

1. relaxed clock log normal, calibrated Yule model, uniform distribution, chain lenght 20'000.000 and trace log 1.000. The results are:

2. relaxed relaxed clock log normal, calibrated Yule model, uniform distribution, chain lenght 50'000.000 and trace log 1.000. The results:

As you can see my ESS results are very low. Does anyone knows how to rise these statistics?

I'm planning on changing calibrated yule to only yule, and my priors as following: on brithrate applying a uniform distribution, ucdl mean on gamma with alpha = 0.001 and beta = 10.000, ucdl stdev exponential with mean 0.5.

Can anyone tell me if this priors are optimal or not?
Thanks in advance

[Jordan Douglas]

Hi Gabriela,

1. It's important to use an informed prior for the clock rate (ucldMean) because it is strongly correlated with the tree height, and this causes a lot of mixing issues (and may also explain why your ucldMean estimates are very different in the two chains. Uniform(0,Inf) can give a lot of issues, and Gamma(0.001, 10,000) is probably also too broad. I suggest you try to find an informed clock rate prior - there is a good guide here

https://beast2-dev.github.io/hmc/hmc/Priors/ClockPrior/

2. The Yule prior birth rate should also have an informed prior. There is a guide for this here

https://beast2-dev.github.io/hmc/hmc/Priors/YuleBirthRatePrior/

3. The relaxed clock model mixes quite slowly, especially the standard deviation (ucldStdev)as you have observed. I suggest you use the ORC package to make this mix faster. This is the same model as the default relaxed clock but it is vastly more efficient, particularly on larger datasets. Instructions can be found here:

https://github.com/jordandouglas/ORC

Hope this helps, 

Jordan

[Farman, Mark L.]

There’s obviously something VERY wrong with your data. You say you’re working with fossils, yet your TMRCA is 72 years!  Your clock rate is also far too high. Have you checked your temporal signal?

Mark L. Farman 
Professor, Department of Plant Pathology

225 Plant Science Building
1405 Veteran's Dr.
University of Kentucky
Lexington, KY 40546 USA 
tel:  (859) 218-0728 
fax:  (859) 323-1961

Sent from my iPhone

On Oct 3, 2022, at 6:15 PM, Gabriela Cruz Luna <est.gabr...@unimilitar.edu.co> wrote:



CAUTION: External Sender

Para ver el aviso de privacidad clic aquí y para el uso de datos personales clic aquí. 

--
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/54906bd8-7883-411f-8f73-9258e7717f00n%40googlegroups.com.

[Jordan Douglas]

Hi Mark,

Disagree - TMRCA can be in any units (eg millions of years)  and thus so can the clock rate. I believe the main issue here is that clock rate has an improper prior (defaults to uniform(0,Inf)). I don't see any reason to conclude that something is wrong with the data from this information alone

Cheers,

Jordan

[Gabriela Cruz Luna]

Thanks everyone for one and each of your responses

Just wanted to tell you that using the parameters I told you (changing calibrated yule to only yule, and my priors as following: on brithrate applying a uniform distribution, ucdl mean on gamma with alpha = 0.001 and beta = 10.000, ucdl stdev exponential with mean 0.5) my results came out like this

So I believe that all my statistics are optimal and I can continue my job!!!!

Hope this information help someone else

To view this discussion on the web visit https://groups.google.com/d/msgid/beast-users/94d785de-273c-4698-86fa-33d4b5a276adn%40googlegroups.com.