Coalescent likelihood error

[Jay Taylor]

Dear BEAST users,

During analyses of a fairly large data set (691 sequences of length
5800) using the constant size coalescent as a tree prior and a strict
molecular clock, I find that BEAST initially gives the warning

SEVERE: CoalescentLikelihood is -Infinity

multiple times before apparently hitting upon an acceptable set of
parameter values and then proceeding through the run. When run for
sufficiently many generations, the posterior distributions of the
substitution rate parameters and the TMRCA all seem to converge to
reasonable values. In
contrast, if I use the Yule process as the tree prior, then BEAST does
not give this warning and the run proceeds normally and also produces
seemingly reasonable estimates. In both cases, I am using the UPGMA
generated tree as the initial tree and there are no topological or
date constraints on the genealogy. Also, for the coalescent-based
analyses, I am using the `estimated initial root height' assigned by
BEAUti (it won't let me change this on the Tree prior panel) as well
as Jeffrey's prior on the coalescent population size with the
`estimated initial root height' as the initial population size.
Oddly, if I try changing either the initial root height or the initial
population size in the XML file, then BEAST still warns that the
CoalescentLikelihood is -Infinity, but then aborts the run.

I am wondering why these two priors behave so differently at the
outset of the runs and whether I should `trust' the results of the
coalescent-based analyses. Also, is there a way to set up the runs so
that the initial likelihood under the coalescent is not negligibly
small?

Thanks for any advice,
Jay

[Andrew Rambaut]

Dear Jay,

This is an odd error for the constant size coalescent. It means the coalescent density for an interval is zero (thus the log is -inf). I suspect it is because there is a zero interval between nodes but it seems odd this doesn't happen at the initial evaluation (in which case BEAST would just stop). As it is, it will simply be rejecting proposals that give this error and as the messages disappear it suggests it is something about the initial tree. I suggest you try not using the UPGMA starting tree and seeing if that fixes the issue (you may need to use the BEAGLE library to be able to calculate the initial tree likelihood with sufficient precision).

Best,
Andrew

> --
> You received this message because you are subscribed to the Google Groups "beast-users" group.
> 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.
>

[Jay Taylor]

Thanks for the suggestion, Andrew. If I use one of the trees
generated in the later stages
of a BEAST run for a starting tree, then the new run proceeds without
generating the
CoalescentLikelihood warning. Incidentally, I was using the BEAGLE
library with double
precision, so that wasn't the source of the problem.

For reference, I have copied the output produced by BEAST when I use
the UPGMA-generated
tree below.

Best wishes,
Jay

--------------------------------------------------------------------------------------------------------------------------------------------------------

state Posterior Prior Likelihood rootHeight
noncoding.clock.rate coding.clock.rate
0 -18590.1204 -5287.3146 -13302.8058 1.17085E-3 1.00000
1.00000 -
Aug 16, 2011 11:35:52 PM dr.evomodel.coalescent.CoalescentLikelihood
calculateLogLikelihood
SEVERE: CoalescentLikelihood is -Infinity
Aug 16, 2011 11:35:52 PM dr.evomodel.coalescent.CoalescentLikelihood
calculateLogLikelihood
SEVERE: CoalescentLikelihood is -Infinity
Aug 16, 2011 11:35:52 PM dr.evomodel.coalescent.CoalescentLikelihood
calculateLogLikelihood
SEVERE: CoalescentLikelihood is -Infinity
Aug 16, 2011 11:35:52 PM dr.evomodel.coalescent.CoalescentLikelihood
calculateLogLikelihood
SEVERE: CoalescentLikelihood is -Infinity
Aug 16, 2011 11:35:52 PM dr.evomodel.coalescent.CoalescentLikelihood
calculateLogLikelihood
SEVERE: CoalescentLikelihood is -Infinity
Aug 16, 2011 11:35:52 PM dr.evomodel.coalescent.CoalescentLikelihood
calculateLogLikelihood
SEVERE: CoalescentLikelihood is -Infinity
Aug 16, 2011 11:36:01 PM dr.evomodel.coalescent.CoalescentLikelihood
calculateLogLikelihood
SEVERE: CoalescentLikelihood is -Infinity
Aug 16, 2011 11:36:01 PM dr.evomodel.coalescent.CoalescentLikelihood
calculateLogLikelihood
SEVERE: CoalescentLikelihood is -Infinity
Aug 16, 2011 11:36:26 PM dr.evomodel.coalescent.CoalescentLikelihood
calculateLogLikelihood
SEVERE: CoalescentLikelihood is -Infinity
Aug 16, 2011 11:36:26 PM dr.evomodel.coalescent.CoalescentLikelihood
calculateLogLikelihood
SEVERE: CoalescentLikelihood is -Infinity
Aug 16, 2011 11:36:31 PM dr.evomodel.coalescent.CoalescentLikelihood
calculateLogLikelihood
SEVERE: CoalescentLikelihood is -Infinity
Aug 16, 2011 11:36:31 PM dr.evomodel.coalescent.CoalescentLikelihood
calculateLogLikelihood
SEVERE: CoalescentLikelihood is -Infinity
Aug 16, 2011 11:36:32 PM dr.evomodel.coalescent.CoalescentLikelihood
calculateLogLikelihood
SEVERE: CoalescentLikelihood is -Infinity
Aug 16, 2011 11:36:32 PM dr.evomodel.coalescent.CoalescentLikelihood
calculateLogLikelihood
SEVERE: CoalescentLikelihood is -Infinity
10000 -11307.2857 1993.9934 -13301.2791 2.5864E-3 1.05864
0.95501 -
20000 -11333.4495 1993.2425 -13326.6920 2.965E-3 1.16798
0.87113 2.04 hours/million states
30000 -11304.2746 2029.8895 -13334.1641 3.0356E-3 1.10747
0.91755 2.04 hours/million states
40000 -11377.4713 1956.9117 -13334.3830 2.52268E-3 1.17256
0.86762 2.02 hours/million states