Strange convergence behaviour: Hairy caterpillar slowly changing values

[Amanda]

Hello :)

I am running BEAST2 and observe behavior like this in many cases. It seems like the steps taken by the program are very small, and ESS values end up being very low, even though I increase chain length to above 100 million states.

Do you know why it has difficulties converging in cases like this?

Thank you in advance,

Best,

Amanda

[Terrence Demos]

The lower Trace file looks fine. The upper one needs to run for longer, try 300 million. And make sure your Burn-in in Tracer is large enough to sample from the stationary traces only

[Alexei Drummond]

Hi Amanda,

This mixing looks pretty terrible. It is hard to diagnose why without knowing something about the model. How many partitions, how many sequences, what models, any node calibrations? Are there other parameters besides the log likelihoods that are not converging. That will probably be the most helpful for diagnosing the problem.

Cheers

Alexei

On 25/05/2021, at 6:04 AM, Amanda <dynamis...@gmail.com> wrote:

Hello :)

I am running BEAST2 and observe behavior like this in many cases. It seems like the steps taken by the program are very small, and ESS values end up being very low, even though I increase chain length to above 100 million states.

Do you know why it has difficulties converging in cases like this?

Thank you in advance,

Best, <Screenshot 2021-05-24 at 20.04.12.png>

Amanda

<Screenshot 2021-05-24 at 11.50.44.png>

--
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/746c570b-d035-4b8e-9c0c-bf4bc49c32f0n%40googlegroups.com.
<Screenshot 2021-05-24 at 11.50.44.png><Screenshot 2021-05-24 at 20.04.12.png>

[Amanda]

Hi,

thank you for your replies.

In this particular case, number of sequences is 36, but I have started observing the same with my 244 seq dataset that is an extension of a dataset usually converging within 50 mio. states with BDSKY and a strict clock. Both datasets are unpartitioned.

The sequences are between 0 and 40 years old, and I do not calibrate the nodes.

I see this strange moving caterpillar for many parameters, e.g. this treeheight, but the tree likelihood, prior, posterior, rate params etc all have similar behaviours:

I have also tried using mascot as tree prior instead of BDSKY with similar results.

Maybe it is just me who is used to misspecified models looking far more like bimodal traces, noisy signals or similar, and not as a slowly rising caterpillar :)

/Amanda

[Alexei Drummond]

Dear Amanda,

When you say you have “started observing” this problem and that it was “usually converging” within 50M, what has changed? Have you changed the version of BEAST, or the model, or just the number of sequences? Do you have an example XML that mixes well, and another that mixes poorly that you could share?

BTW: A well mixed trace will always look like a horizontal hairy caterpillar. It should not be slowly rising. If it is slowly rising then it hasn’t reached convergence yet.

Cheers

Alexei

On 25/05/2021, at 7:44 PM, Amanda <dynamis...@gmail.com> wrote:

Hi,

thank you for your replies.

In this particular case, number of sequences is 36, but I have started observing the same with my 244 seq dataset that is an extension of a dataset usually converging within 50 mio. states with BDSKY and a strict clock. Both datasets are unpartitioned.

The sequences are between 0 and 40 years old, and I do not calibrate the nodes.

I see this strange moving caterpillar for many parameters, e.g. this treeheight, but the tree likelihood, prior, posterior, rate params etc all have similar behaviours:

To view this discussion on the web visit https://groups.google.com/d/msgid/beast-users/2c334c07-0825-4713-9bc6-52f75f87a4bdn%40googlegroups.com.
<Screenshot 2021-05-25 at 09.31.40.png>