BSP Not converging

[Abid Ali]

Dear Beast Community,

I am running in BSP analysis in BEAST version 1.10.4 using CO1 and CYTB concatenated datasets to explore clade-specific demography. Despite extensive efforts, which involved running billions of chains and testing variations in the number of groups (ranging from 4 to 15 consecutively), I have been unable to obtain ESS.

For the clock rate, I utilized a value of 0.0476, derived from a time divergence tree using a constant coalescent population model. Additionally, I applied a strict clock, and the evolutionary model, detailed below, remains consistent with the methodology used for the entire dataset to estimate the MRCA under the constant coalescent population model suggested by Partitionfinder. However, none of the four clades show convergence.

 Please, see the information below,

I am looking forward your kind suggestion and help in this regard, thank you!

Abid

[HS]

Hi,

One thing that may help to improve ESS, might be using Gamma instead of Gamma+Invariants. You may find explanations in this group of why the Gamma+Invariants model might be problematic.

Best,

Hovhannes

[Abid Ali]

Dear Havhannes,

Thank you so much for your insight. Well, I tried the one you are suggesting and still failed to get convergence on the data. I then ran Partitionfinder on each corresponding lineage dataset to find a model, however, still did not work after having a separate model.

I'm wondering what might help to increase the ESS value.

Thanks

--
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/317a6d90-9071-4155-b6dc-aac949a9caddn%40googlegroups.com.

[Artem B]

Hello Abid,

ESS of, for example, 60 means that you should increase the length of MCMC by 4 to get ESS about 240. Depending on a data set, I have to run 100-1,000 millions + iterations from time to time, it's OK.

Best,
Artem

среда, 7 февраля 2024 г. в 14:23:20 UTC+8, Abid Ali:

[jinlin shen]

Dear Community Members,

I hope this message finds you well. I am reaching out for some insights and assistance regarding an interesting phenomenon I encountered while using Tracer to assess the performance of my Bayesian analysis.

Background: I ran a Bayesian analysis with a chain length of 300 million. Following the community's advice, I used gamma distributions to replace the Gamma+Invariants in my model.

Observation: Initially, at around 160 million iterations, I observed that my Effective Sample Size (ESS) values for the joint, prior, and posterior distributions were around 56. This seemed reasonable and gave me confidence in the analysis.

However, to my surprise, upon completion of the full 300 million iterations, the ESS values for these same parameters dropped significantly to only around 10.

Question: I am curious to understand what might cause such a drastic change in ESS values towards the end of the run. Are there any common pitfalls or factors that I might have overlooked?

Request for Help: Any guidance or suggestions on how to interpret this drop in ESS or steps to diagnose the issue would be greatly appreciated. I am keen to ensure that my analysis is robust and the results are reliable.

Thank you in advance for your time and expertise. I am looking forward to any insights the community can provide.

Best regards,

[Alexei Drummond]

Hi Jinlin Shen,

Did you look at the traces in Tracer? Are there sudden jumps in the trace and if so in which direction? Or are there long slow wavelengths? Visual inspection of the trace is one of the steps recommended in order to evaluate whether the MCMC run is mixing or not. It is often easier to diagnose what is happening when both the ESS and the trace are available. Roughly speaking you would like to be able to cut the trace in half and see that the first and the second half look the same (recursively for some number of steps).

Cheers

Alexei

--
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/10da143d-8c03-458f-b6fb-6a83aba65017n%40googlegroups.com.

[jinlin shen]

Hi Alexel,

I apologize for seeing your reply so late. Thank you very much. As you said, according to my understanding, there are indeed sudden ups and downs here, but how should this be resolved? I look forward to your reply.

Cheers

Jinlin Shen

[jinlin shen]

Oh, by the way, here is the screenshot of my TRACER analysis. I hope it helps you understand my difficulties more clearly."

在2024年5月17日星期五 UTC+8 05:25:02<Alexei Drummond> 写道：