Hi All,
Sorry to keep bringing this up, but I'm trying to satisfy some reviewers. I have a rapid event
related design. The criticism that has been made is that because the hemodynamic response from
each event is not completely orthogonal, it is impossible to resolve the individual events. I am
trying to quantify the degree of orthogonality in my model and demonstrate that the design
matrix is not rank deficient.
That's why I was hoping to find a condition number threshold commonly used in linear regression
or a rationale for the threshold used in FSL.
Alternatively, does anyone have any suggestions for other methods to do the same thing.
thanks,
jack
On Sat, 23 Oct 2004 20:12:14 +0100, Jack Grinband <[log in to unmask]> wrote:
>Hi Steve,
>Just wanted to follow up:
>
>For which design types is this test appropriate/inappropriate?
>
>I've found numerous references to the condition number, but none of them have stated any
actual
>threshold values that are commonly used. How did you arrive at 5e-2? Have you tested
>simulated data to see how low you could go before the GLM failed?
>
>many thanks,
>
>jack
>
>
>On Wed, 20 Oct 2004 09:12:23 +0100, Stephen Smith <[log in to unmask]> wrote:
>
>>Hi Jack. Yes, the program feat_model (which creates the model given the
>>design, including outputting the ratio _only if_ it faile test) could be
>>recompiled to always output this - you just need to find the relevant
>>printf statement and recompile....
>>
>>The threshold is currently set at 5e-2 - the reason that we have talked
>>about different thresholds is that doing the test the way we do (via the
>>ratoi of min to max eigenvalues) isn't appropriate for all design types -
>>basically it's just more complicated than that. However, we're just now
>>working on a better test which takes into account the contrasts and
>>typical noise level etc - so watch this space.
>>
>>Cheers.
>>
>>On Wed, 20 Oct 2004, Jack Grinband wrote:
>>
>>> Hi All,
>>> What exactly is FEAT's rank deficiency threshold? I searched through the list and various
>replies
>>> have listed e(-8), e(-6), e(-5), and e(-4) as "conservative" thresholds. There is 4 orders of
>>> magnitude between the smallest and largest of these. So, does that mean that the threshold
>>> depends on some other factors?
>>>
>>> Also, I'm having trouble finding the eigenvalue min:max ratio for each analysis. I couldn't
find
>it in
>>> the log or design files. Is there any way I can get this information?
>>> thanks,
>>>
>>> jack
>>>
>>
>> Stephen M. Smith DPhil
>> Associate Director, FMRIB and Analysis Research Coordinator
>>
>> Oxford University Centre for Functional MRI of the Brain
>> John Radcliffe Hospital, Headington, Oxford OX3 9DU, UK
>> +44 (0) 1865 222726 (fax 222717)
>>
>> [log in to unmask] http://www.fmrib.ox.ac.uk/~steve
|