Welcome, Guest. Please Login
YaBB - Yet another Bulletin Board
 
  HomeHelpSearchLogin  
 
Std. dev. values are all identical (Read 1419 times)
Jeff Mower
Guest


Std. dev. values are all identical
Nov 30th, 2004 at 10:14am
 
Hello,

I am using the windows version of hyphy and have two data sets of about 50 taxa and around 1500 characters.  I have used both the GY94w9 and MG94w9 codon models with local/independent parameters to optimize the likelihood for a given topology.  After the ML optimization step, which seems to complete fine, I tell the program to calculate variances.  Regardless of codon model, data set, or variance calculation method (finer vs. crude), the standard deviations reported are identical for all parameters.  The CI also has no range, i.e., the min and max are the same values as the parameter itself.  The parameters themselves are different from each other, however.  Interestingly, when I calculate CI using the likelihood profile, it seems to work fine.


I took one of these data sets and reduced it to only 8 taxa.  Now using this data set and setting up the analyses exactly as above, the variances seem to calculate properly regardless of codon model or calculation method.


Any suggestions?  Thanks for your help.

Jeff
Back to top
 
 
IP Logged
 
Sergei
YaBB Administrator
*****
Offline


Datamonkeys are forever...

Posts: 1658
UCSD
Gender: male
Re: Std. dev. values are all identical
Reply #1 - Dec 1st, 2004 at 1:58am
 
Dear Jeff,

I am almost sure that what is happening on the 50 sequence data set is that the Fisher Information Matrix is singular or not positive-definite (this is possible because derivatives are computed numerically). If you could kindly send me your data (spond@ucsd.edu) I'd like to check that this is the case, and if it is, I'll update the code to produce an error message rather than nonsensical CI.

Likelihood profile approach does not suffer from the singularity issue, and in general is preferable to methods based on asymptotic normality (because we can't really be sure that the likelihood surface is almost normal for finite size data sets).

Hope this helps,
Sergei
Back to top
 

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