low ESS values for prior and posterior - revisited

[Chris Law]

Dear all,

I'm running a 46 gene partition dataset with 77 taxa and 11 fossil calibrations. After increasing my generation time from 10,000,000 to 100,000,000, I still get very low ESS scores (~10) for my prior and posterior parameters; all other ESS values are very high (> 1000). I've tried the suggestions from previous posts about this topic but my ESS values remain low. I was wondering if you had any advice on how to improve the prior and posterior ESS scores (xml file attached). 

I appreciate your help.

Thanks!

Chris

[Remco Bouckaert]

Dear Chris,

Your analysis contains quite a lot of parameters as well as a number of calibrations, which may be the cause of the bad mixing you experience. I would try to reduce the number of parameters by using simpler substitution models, starting with HKY with empirical frequencies and see whether that gives decent ESSs. If so, you can add more parameters/use more complex models.

Since some of the models are GTR and other HKY at the moment, I assume you used something like JModelTest to choose these models. An alternative is to use the RB  model (you need to install the RBS package), which uses reversible jump to choose the substitution model and jointly estimates the appropriate number of parameters while sampling the tree (instead of picking models before the analysis).

Cheers,

Remco

[Chris Law]

Thanks Remco for the suggestion with the RBS package.

Just to make sure I'm using it correctly ) set the substitution model as the RB model. Do I leave substitution rate as 1, gamma category count as 0, and proportion invariant as 0?

Thanks again!

Chris Law

--
You received this message because you are subscribed to a topic in the Google Groups "beast-users" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/beast-users/doUyUxXnizM/unsubscribe.
To unsubscribe from this group and all its topics, send an email to beast-users...@googlegroups.com.
To post to this group, send email to beast...@googlegroups.com.
Visit this group at http://groups.google.com/group/beast-users.
For more options, visit https://groups.google.com/d/optout.

[Ashok Mallik]

Hello Chris, 

I am also getting same problem with low ESS. Are you able to use RB model? What is the result?

Thanks

Ashok

[Remco Bouckaert]

Dear Chris,

Gamma rate heterogeneity and proportion invariant are not estimated by the RB-model, only the substitution model. Since you had trouble getting convergence, I would start with the settings you suggested, and once this analysis mixes, try adding rate heterogeneity (set gamma category count to 4, and check estimate next to the shape parameter).  In either case, you can estimate the proportion invariant for those partitions that have a significant number of constant sites, but leave the proportion invariant at zero for those partitions that are all variable.

Cheers, 

Remco

[Chris Law]

Hi Remco,

I still receive low ESS numbers when running my analysis with the RB model at 150,000,000 generations. I am trying it at 500,000,000 generations now, but I want to try to use a starting tree (fr/ MrBayes) to run my Beast analysis faster. 

I replaced the RandomTree state-node initialiser with the following:

<init spec="beast.util.TreeParser" id="NewickTree.t:musteloidtree"

  initial="@Tree.t:tree" taxa="@ADORA3" IsLabelledNewick="true" 

  newick="(Ursus_arctos:1,Phoca_vitulina:1,(((((((((((((((Aonyx_capensis:0.09254622127, AND REST OF TREE)

/>

However, after following the instructions on the Beast blog, I get this error:

java.lang.Exception: Could not find a proper state to initialise. Perhaps try another seed.

at beast.core.MCMC.run(Unknown Source)

at beast.app.BeastMCMC.run(Unknown Source)

at beast.app.beastapp.BeastMain.<init>(Unknown Source)

at beast.app.beastapp.BeastMain.main(Unknown Source)

at sun.reflect.NativeMethodAccessorImpl.invoke0(Native Method)

at sun.reflect.NativeMethodAccessorImpl.invoke(NativeMethodAccessorImpl.java:39)

at sun.reflect.DelegatingMethodAccessorImpl.invoke(DelegatingMethodAccessorImpl.java:25)

at java.lang.reflect.Method.invoke(Method.java:597)

at apple.launcher.LaunchRunner.run(LaunchRunner.java:116)

at apple.launcher.LaunchRunner.callMain(LaunchRunner.java:51)

at apple.launcher.JavaApplicationLauncher.launch(JavaApplicationLauncher.java:52)

Do you have any suggestions as to how I can fix this?

I appreciate your help.

Chris

Chris Law

[Remco Bouckaert]

Hi Chris,

Thanks for the file. The reason it does not start is because the calibrations are incompatible with the tree — not because of the monophyletic constraints, but because the tree is too low. There are log-normal distributions that have an offset that is too high for the tree to fit.

To fix this, change the offset to 0 and set meanInRealSpace=true. Then change the mean of the lognormal distribution to be the mean of the age for that clade. The only thing left is then to set the variance to a small value such that the 90% or 95% HPD fits that of the one you had in mind — the most convenient way is to do all this in BEAUti and try a few values in the Priors panel. The panel will show the 90% and 95% HPD intervals.

By setting up the log normal distribution this way, you guarantee that there is some support for small trees, but the majority of the probability mass is around the mean, so that is where the tree will be drawn to. 

It is very well possible that your calibrations and the data do not fit well together, which causes poor convergence. You can verify this by leaving one calibration (say the oldest) in place, and remove all distributions from the other calibrations but leave the monophyletic constraints in place (set distribution to ’none’ in BEAUti). If you run the analysis and get good convergence, you can have a look at where the data puts the clades and compare them to your calibrations.

Cheers,

Remco

[fjfl...@espe.edu.ec]

Make sure that dates are set correctly. Here is an example of what worked for me