HyPhy message board - Log-likelihood in PARRIS analysis

	Welcome, Guest. Please Login

Home

Help

HyPhy message board › Methodology Questions › How to › Log-likelihood in PARRIS analysis

(Moderators: Sergei, Simon)

‹ Previous Topic | Next Topic ›

Pages: 1

Send Topic

Log-likelihood in PARRIS analysis (Read 2026 times)

CGS

YaBB Newbies

Offline

Feed your monkey!

Posts: 8

Log-likelihood in PARRIS analysis
Dec 19^th, 2007 at 11:36am

Hi all,
I have run the PARRIS analysis in HyPhy and I got strange results (below).. Is it possible that the log-likelihood of the M2a model be lower than the M1a model?

RUNNING MG94012212 MODEL COMPARISONS on 0
########### 3x3 CLASSES ###########
+--------------------+----------------+---------------+-----------------+-------
-----------+-----------+-----+-----------+
| Model | Log Likelihood | Synonymous CV | NS Exp and CV | N/S Exp and CV | P-Value | Prm | AIC |
+--------------------+----------------+---------------+-----------------+-------
-----------+-----------+-----+-----------+
| discr(3), M1a | -3074.88984 | 0.98100808 | 0.30413,1.46539 | 0.66934, 1.98894 | N/A | 145| 6439.78|
+--------------------+----------------+---------------+-----------------+-------
-----------+-----------+-----+-----------+
| discr(3), M2a | -3078.14338 | 0.84220784 | 0.36350,1.65439 | 0.53952, 1.81602 | 1.0000000 | 148| 6452.29|
+--------------------+----------------+---------------+-----------------+-------
-----------+-----------+-----+-----------+

IP Logged

Sergei

YaBB Administrator

Offline

Datamonkeys are forever...

Posts: 1658
UCSD
Gender: male

Re: Log-likelihood in PARRIS analysis
Reply #1 - Dec 20^th, 2007 at 8:17am

Dear CGS,

The LogL for M2a should not be less than than of LogL for M1a; what you are seeing is a convergence problem.
Try running the analysis a few times with random starting values (there should be a dialog option for that) to see how the LogL behaves. Convergence problems indicate some sort of a statistical pathology and tend to happen with smaller and/or low divergence alignments when one of the models is overfitting the data. Another thing you may try is changing optimization precision. In GUI versions you can use the preferences dialog, and in command line version make a .hyphyinit file in HYPHY base/working directory and add a line
Code:

OPTIMIZATION_PRECISION = 0.0001; /* adjust as needed; default is 0.001 */

I don't think that will help that much though, as the default (0.001) is nearly always adequate.

Cheers,
Sergei

Associate Professor
Division of Infectious Diseases
Division of Biomedical Informatics
School of Medicine
University of California San Diego

WWW

IP Logged

CGS

YaBB Newbies

Offline

Feed your monkey!

Posts: 8

Re: Log-likelihood in PARRIS analysis
Reply #2 - Dec 21^st, 2007 at 10:46am

I have set the optimization precision to 0.00000001 and it seems to work fine now. Here is the new result. Thanks Sergei.

RUNNING MG94012212 MODEL COMPARISONS on 0
########### 3x3 CLASSES ###########
+--------------------+----------------+---------------+-----------------+-------
-----------+-----------+-----+-----------+
| Model | Log Likelihood | Synonymous CV | NS Exp and CV | N/S Exp and CV | P-Value | Prm | AIC |
+--------------------+----------------+---------------+-----------------+-------
-----------+-----------+-----+-----------+
| discr(3), M1a | -3078.55725 | 0.99329655 | 0.30381,1.46639 | 0.63275, 1.93029 | N/A | 148| 6453.11|
+--------------------+----------------+---------------+-----------------+-------
-----------+-----------+-----+-----------+
| discr(3), M2a | -3070.78986 | 0.99696987 | 0.36381,1.65875 | 0.74427, 2.10368 | 0.0004233 | 150| 6441.58|
+--------------------+----------------+---------------+-----------------+-------
-----------+-----------+-----+-----------+

IP Logged

Sergei YaBB Administrator Offline Datamonkeys are forever... Posts: 1658 UCSD Gender:	Re: Log-likelihood in PARRIS analysis Reply #3 - Dec 21^st, 2007 at 12:41pm Dear GCS, Glad it worked out! These models (M1a M2a etc) with extra twists (like syn. rate variations) are sometimes numerically unstable, unfortunately! Cheers, Sergei
Back to top	Associate Professor Division of Infectious Diseases Division of Biomedical Informatics School of Medicine University of California San Diego WWW IP Logged

Pages: 1

Send Topic

‹ Previous Topic | Next Topic ›

« Home

‹ Board

Top of this page