Bad convergence, likelihood going down (STACEY, BEAST2)

mariheilertsen

unread,

Dec 22, 2016, 3:20:52 AM12/22/16

to beast-users

Hi all,

I am running a STACEY-analysis on a dataset with four markers and quite a bit of missing data. Each specimen has at least two markers, and all specimens are set as one species. Looking at the log-file in tracer after 80m generations the analysis seems to be running towards worse likelihoods, this can not be right? ESS values are not good, and not improving much with time, so I do not think it will help to just run it longer either. Could it be bad priors? Over-parameterized model? Or because of missing data?

Hoping for some input.

Thanks,

Mari

Screen Shot 2016-12-22 at 09.07.00.png

Graham

unread,

Dec 22, 2016, 7:33:15 AM12/22/16

to beast-users

Decreasing likelihoods does not necessarily mean something is wrong. You may just need to run it for much longer. It might be useful to try with a small number of individuals to check that it converges in that case.

Graham Jones

mariheilertsen

unread,

Dec 22, 2016, 8:25:41 AM12/22/16

to beast-users

Thank you for the quick reply.

If I run a test with the same priors and just a subset of the species, and that analysis converges, does that mean I just need to run it longer?

At 100m generations the likelihood has ok ESS values (just over 200), but the prior and posterior has an ESS just over 5. Tree height and mutation rate parameters are also in that range (ESS 5-6) How long do you think I need to run it? 500m generations?

Mari

Graham

unread,

Dec 22, 2016, 9:08:45 AM12/22/16

to beast-users

On Thursday, 22 December 2016 13:25:41 UTC, mariheilertsen wrote:

Thank you for the quick reply.

If I run a test with the same priors and just a subset of the species, and that analysis converges, does that mean I just need to run it longer?

Probably, but it's not a guarantee that nothing is wrong.

At 100m generations the likelihood has ok ESS values (just over 200), but the prior and posterior has an ESS just over 5. Tree height and mutation rate parameters are also in that range (ESS 5-6) How long do you think I need to run it? 500m generations?

I don't know, but probably longer than that.

mariheilertsen

unread,

Dec 22, 2016, 12:32:59 PM12/22/16

to beast-users

I tried to follow the suggestions in the Stacey manual for priors, do you have any suggestions on ways to simplify the model to improve convergence?

Graham

unread,

Dec 23, 2016, 3:37:13 AM12/23/16

to beast-users

Maybe: strict clocks, no site rate heterogeneity, HKY substitution model.

mariheilertsen

unread,

Dec 23, 2016, 2:11:32 PM12/23/16

to beast-users

Thank you very much for your input, I tried to run an analysis with less than half of the dataset, and it did not converge well within 100m generations either, so I will try out your suggestions to simplify the model.

Mari

Reply all

Reply to author

Forward