low ESS values for "prior" and "posterior"

[Ted]

I have a question very similar to one that was posted several months
ago by someone else but didn't receive any responses. I ran two
separate identical analyses for a 4-gene, 12 partition (by codon
partition), 109 taxa data set, with several calibration nodes
specified. Ran them each for 30,000,000 generations. Looking at the
results in Tracer, the likelihood values for both runs seemed to
quickly reach the same range and remain there for the duration of the
runs. ESSs for most all parameters for all partitions were at least
in the black (>100) and most were very high (the likelihood ESS was
~1800). However, the "prior" and "posterior" ESSs were still very low
(~13 for each for the combined runs), and the traces for these two
showed them moving around quite a bit. What exactly are these two
values telling me, and if all my other ESSs are OK, do these low
values necessarily mean my results are not reliable (the consensus
tree I'm getting from these analyses looks reasonable, i.e., basically
identical to that from an unrooted MrBayes analysis, and the PPs are
very high on almost all nodes)?

Thanks for any ideas,

Ted

[Andrew Rambaut]

Dear Ted,

I suspect this may be the same problem I have seen before - Are you
using the GTR model on all (or any) of your partitions? If so, check
if one of the relative rates is going to zero. When this happens, the
Jeffrey's prior on the rate becomes invalid. The rate may be going to
zero because there aren't any of that type of substitution observered.

The solution is either to use the HKY model or try removing the
Jeffrey's
priors from the GTR rates.

If that is not the problem, get back to me.

Andrew

___________________________________________________________________
Andrew Rambaut
Institute of Evolutionary Biology University of Edinburgh
Ashworth Laboratories Edinburgh EH9 3JT
EMAIL - a.ra...@ed.ac.uk TEL - +44 131 6508624

[tyl...@hkucc.hku.hk]

Dear Andrew and Ted,

I also the similar situation before (I used GTR). And I found changing the
"Jeffrey's" prior back to "uniform" prior could improve the ESS of prior and
posterior in my case.

Best wishes,
Tommy

[Suzanne Williams]

Dear Andrew and Ted,

I changed the priors from "Jeffrey's" to "lognormal" and this also fixed the
problem.

Thanks for the hint!

Suzanne

----------------------------------------------------------
A one day symposium on "Speciation in Molluscs" will be held at the NHM,
London on 25th April 2008. Speakers are Angus Davison, Chris Meyer, Ellinor
Michel, Emilio Rolán-Alvarez, Menno Schilthuizen and Jon Todd.

Registration is FREE. Please email s.wil...@nhm.ac.uk if you wish to
attend.

--------------------------------------------------------
Dr Suzanne Williams
Zoology Dept
Natural History Museum
Cromwell Rd
London SW7 5BD
United Kingdom
Tel: + 44 (0) 207 942 5351 (office) 5774 (lab)
Fax: +44 (0) 207 942 5867

http://www.nhm.ac.uk//research-curation/staff-directory/zoology/cv-5597.html

[danrabosky]

Dear all -

Are there any strong reasons in favor of using Jeffrey's prior over a
uniform or log-uniform for relative rates with GTR? I am happy to have
seen these posts - I was finding that my priors were blowing up as
several relative rates approached zero; things are now working great &
my ESS is much improved. To avoid this problem from the start in
future runs, would it be advisable to simply not use Jeffrey's prior?
Has anyone performed any comparisons of distributions of topologies /
branch lengths with and without Jeffrey's prior?

Cheers,
Dan

[Ted Townsend]

Hi Dan,

I'm sure you'll receive better technical explanations of what's happening, but with regards to your last question, I was one of the ones who was asking recently, and I reran my analysis twice with uniform priors (the default if nothing is specified in the BEAST file) per Andrew's suggestion. Everything came out great ESS-wise with those analyses, the topology was the same as before, and the tmrcas were across the board almost exactly what they were with the Jeffreys analyses. My impression is that using the Jeffreys priors didn't change anything about the results, but just made it impossible to know whether there was really any problem with the search or whether the low ESS for the posterior was just an artifact.

Ted

[alexei....@gmail.com]

Hey Dan,

This problem with the Jeffreys prior only occurs when you are using a
model that is basically too parameter-rich for your data. The Jeffreys
prior provides better statistical properties for estimating the kappa
parameter in the HKY model and the relative rates parameters in GTR
*provided* that the data analyzes actually has a few events of each
type of mutation in it (see Drummond et al 2002, Table 3).

However, in my experience, if you don't care about accurately
estimating the relative rates and your main focus is on the divergence
times and tree topology, then the choice between Jeffreys and uniform
priors on the relative rates is not a big one. However if Jeffreys
prior is giving you a problem then probably GTR is slightly to complex
a model for your data...

Cheers
Alexei

[石勇]

Dear Alexei,

       I have encounter the similar situation. The ESS of prior distribution is pretty low. 

       I used GTR model for some of the markers . 

       Strangely, I found changing the "Jeffrey's" prior (birthRate.t and popMean) back to "lognormal" prior could not improve the ESS of prior  in my case.

       Could you mind give me some advices? Thank you in advance.

Best wishes,

Sandy

在 2008年1月14日星期一UTC+8上午5时54分11秒，Alexei Drummond写道：

[Michael Forthman]

Hello,

I have encountered the same issues and have attempted to incorporate the solutions given here. I've encountered a slightly different problem as a result. Prior to making the changes, I had inserted a starting tree, fixed the topology, ran BEAST, monitored the .log file (where I noticed low ESS posterior and prior values), and also noted a tree file output. Since making the changes, I do not get any tree file output. I'm assuming that a tree file should still be produced, but I'm not sure why BEAST is no longer doing so after changing the 1/x prior to uniform and/or changing my GTR models to HKY.

[Michael Forthman]

Hello,

I apologize if appears twice (saw an error message when trying to post the first time). I have also encounter low ESS posterior and prior values, and have attempted the solutions given here. Prior to making changes to my xml file, I was performing BEAST on a fixed topology and was receiving both log and tre files. I have since made changes (Jeffrey's prior changed to uniform and/or GTR models changed to HKY). I am no longer getting tre files and not sure why this is happening.

Cheers, Michael

On Sunday, January 13, 2008 1:54:11 PM UTC-8, Alexei Drummond wrote:

[Cinnamon Mittan]

Hi Sandy,

This was from a while ago, but did you ever find a solution to the low prior ESS? I am having the same issue, where all my other values are high, but the ESS for the prior is extremely low. I've also tried switching my models from GTR to HKY, and it actually made the ESS for the posterior much worse. Any advice would be much appreciated!

Best,

Cinnamon