New subject: Problem with glmfit: matrix is ill-conditioned

26 Feb 2013

      Hi Doug,
The condition numbers were:
- 1e+08 for problematic fsgd file
- 371.233 for previous fsgd file
I tried also v5.1, however with no succes. The matirx was still
ill-conditioned with condition number 2.80809e+07.
Cheers,
Ed
On 1 Feb 2013, at 9:53, freesurfer-request@nmr.mgh.harvard.edu wrote:
...
Message: 12
Date: Thu, 31 Jan 2013 14:08:11 -0500
From: Douglas N Greve greve@nmr.mgh.harvard.edu
Subject: Re: [Freesurfer] Problem with glmfit: matrix is
        ill-conditioned
To: freesurfer@nmr.mgh.harvard.edu
Message-ID: 510AC11B.4090902@nmr.mgh.harvard.edu
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Hi Ed, that sounds strange. What were the condition numbers for the  
two
analyses? This should be printed out into the terminal. It is also
possible that it is a scaling issue. This should not be an issue in  
5.1,
but it is possible that it could have an effect if two columns are  
very
similar.
doug
On 01/30/2013 05:10 AM, Ed Gronenschild wrote:
...
Hi Doug,
It is indeed correct that columns 1 and 17 are highly correlated.  
However,
for another version of the fsgd file, where the corresponding  
class and
covariate values where identical but with other covariates (so  
leading to
the same columns  1 and 17, other columns are different), all went  
well.
So probably some other columns are highly correlated as well.
Anyway, what can we do about it? Just leave the offending covariate 
(s)
out?
Cheers,
Ed
...
Message: 1
Date: Thu, 24 Jan 2013 12:51:23 -0500
From: Douglas N Greve <greve@nmr.mgh.harvard.edu
mailto:greve@nmr.mgh.harvard.edu>
Subject: Re: [Freesurfer] Problem with glmfit: matrix is
      ill-conditioned
To: freesurfer@nmr.mgh.harvard.edu
mailto:freesurfer@nmr.mgh.harvard.edu
Message-ID: <5101749B.4060305@nmr.mgh.harvard.edu
mailto:5101749B.4060305@nmr.mgh.harvard.edu>
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Hi Ed, the problem is that your covariates are highly correlated.  
Eg,
column 1 is nearly identical to column 17.
doug
On 01/24/2013 06:36 AM, Ed Gronenschild wrote:
...
Hi,
Running mri_glmfit ( v5.0.0, Mac OSX10.6) I got he following
error message:
ERROR: matrix is ill-conditioned or badly scaled, condno = 1e+08
Possible problem with experimental design:
Check for duplicate entries and/or lack of range of
continuous variables within a class.
If you seek help with this problem, make sure to send:

Your command line:

mri_glmfit --y lh.sib_con_cann_aangepast3.thickness.sm0.mgh  

--fsgd
../../STATS/FSGD/age_gender_scan_educ_hand_cannabis/ 
fsgd_sib_con_cann_aangepast3.fsgd
dods --glmdir lh.sib_con_cann_aangepast3.dods.sm0.glmdir --surf
fsaverage lh --C
../../STATS/FSGD/age_gender_scan_educ_hand_cannabis/ 
group_thickness_gender_scan_age_educ_hand_cann.mtx

The FSGD file (if using one)
And the design matrix above

Please find attached the requested files.
I noticed that rows 68 and 147 of the matrix Xg.dat are identical,
and most probably caused the error.
In the fsgd file I see that the correspondig cases indeed have
identical values, although they are referring to different data.
How to proceed?
Cheers,
Ed

Freesurfer mailing list
Freesurfer@nmr.mgh.harvard.edu
https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
--
Douglas N. Greve, Ph.D.
MGH-NMR Center
greve@nmr.mgh.harvard.edu
Phone Number: 617-724-2358
Fax: 617-726-7422
Bugs: surfer.nmr.mgh.harvard.edu/fswiki/BugReporting
FileDrop: www.nmr.mgh.harvard.edu/facility/filedrop/index.html
Outgoing: ftp://surfer.nmr.mgh.harvard.edu/transfer/outgoing/flat/ 
greve/

Re: [Freesurfer] Problem with glmfit: matrix is ill-conditioned