Model choice and poor convergence

207 views
Skip to first unread message

gbennett

unread,
Jun 20, 2012, 2:27:52 PM6/20/12
to beast...@googlegroups.com
Dear Beast-Users,

I have a methodological/philosophical question regarding model choice and convergence. 

I'm currently running a rather large analyses with for 200 taxa for 6 genes with 14 gene partitions. I have used Modeltest to estimate substitution models for each partition. After running the analyses for different lengths (>100 mil x 8 runs) all but two partitions reach convergence. I specified a GTR for these two partitions and the particular elements that fail to converge include the .ag, .cg, .freq. pinv, etc., receiving ESS between 100-200, or several <100. Everything else converges for independent and combined runs - same tree topology, same divergence dates, same everything.

My question is whether or not it is better to purposely miss-specify a simpler model (e.g. an HKY over a GTR) to reach convergence. Or, maintain the empirically determined, parameter rich model even though convergence is low (...the website suggests convergence on everything may not be necessary). Preliminary runs under simpler models change divergence estimates by few hundred thousand years (the is a Hawaiian system, so that's not totally trivial). With this result, I'm not sure what the more prudent choice is.

I have tried playing around with different priors on the GTR, which hasn't changed much over previous runs (or fails to initiate all together).

Any insights are greatly appreciated!

thanks,
-Gordon

C.-j. Mei

unread,
Jun 20, 2012, 3:32:06 PM6/20/12
to beast...@googlegroups.com
I don't know what the detail is for your system (No. of sequences each group, etc.). But we have a system at ASU which can take a set of sequences, align them, model test them and build the XML based on the model test result and finish the BEAST phylogeography computation from beginning to end. For our computations (as many as 1000 H1N1-HA-NA sequences) all converged.
 
C.-J.

--
You received this message because you are subscribed to the Google Groups "beast-users" group.
To view this discussion on the web visit https://groups.google.com/d/msg/beast-users/-/pS18ij-rZpwJ.
To post to this group, send email to beast...@googlegroups.com.
To unsubscribe from this group, send email to beast-users...@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/beast-users?hl=en.

Reply all
Reply to author
Forward
0 new messages