Dear All,
In my knowledge, in order to do Pearson's correlation, one would need
linearity and joint normality assumptions. For two ranged variables,
e.g. they are not normally distributed in principle and there is only
one set of parameters to establish linear relationship between them over
the whole ranges, and these parameters can be calculated using the range
values rather than the data we acquired. In order to overcome this
situation, two solutions may be possible:
1. Use transformation like logit as Ged pointed out transform the range
of both variable to (-inf inf), and then do the analysis accordingly,
but the interpretation will go wild
2. Leave those variables alone and assume they are "LOCALLY" linearly
related, interpret results with respect to the data ranges encountered
and restrict any extrapolation. Residues should be carefully examined to
verify the assumption. I think this how people deal with most variables
like age.
Best,
Deqiang
-----Original Message-----
From: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]]
On Behalf Of Ged Ridgway
Sent: 2006Äê8ÔÂ1ÈÕ 19:22
To: [log in to unmask]
Subject: Re: [SPM] In need of DTI help
Hi all,
Given that I know nothing about DTI and little about stats, I feel
it's only sensible for me to say something ;-)
If FA is a fractional value between 0 and 1, perhaps a logit transform
would be appropriate? In short, ln(1/(1-FA)) will map (0,1) to
(-inf,inf) and may be more appropriate for correlating with cognitive
scores on a linear scale. (Though I guess the cognitive scores might
not be too linear anyway...)
http://www.ats.ucla.edu/STAT/stata/faq/proportion.htm
http://www.stata.com/support/faqs/stat/logit.html
Regarding the use of percent-range scores, I don't think that a simple
linear transformation of the score (e.g. 100*score/max(score)) will
actually change the correlation coefficient.
Perhaps the best bet would be a non-parametric rank correlation
between the raw FA values and the raw cognitive scores? Better still,
ignore me, and ask a proper statistician!
Best,
Ged.
Volkmar Glauche wrote:
> Dear Deqiang Qiu,
>
> I was a bit short and unclear in my answer yesterday: Of course FA is
an
> important measure for fibre tract integrity and changes in cognitive
> function may be related to changes in white matter structure revealed
in
> FA measurements. But for a valid correlation including significance
> testing, you have to be sure that your data meets some assumptions
(for a
> glossary, see e.g.
http://www.stats.gla.ac.uk/steps/glossary/paired_data.html).
>
> Yours,
>
> Volkmar
>
> On Tue, 1 Aug 2006, dqqiu wrote:
>
>> Dear Volkmar,
>>
>> I'm not quite agree with you in that " Correlating this with
arbitrary
>> unlimited scores will not give meaningful results". I think since FA
>> changes can be indicative of fiber maturation process, including
>> myelination. Correlation of FA and cognitive scores could support
the
>> importance of intact fiber for the proper operation of cognitive
>> function.
>>
>> Best Wishes!
>> Sincerely yours,
>> Deqiang Qiu
>> Department of Diagnostic Radiology
>> The University of Hong Kong
>> Rm 406,Blk K,
>> Queen Mary Hospital
>> Pokfulam Road,
>> Hong Kong
>> Tel: (+852) 28553307
>> Fax: (+852) 28551652
>>
>>
>> -----Original Message-----
>> From: SPM (Statistical Parametric Mapping)
[mailto:[log in to unmask]]
>> On Behalf Of Volkmar Glauche
>> Sent: 2006?8?1? 0:34
>> To: [log in to unmask]
>> Subject: Re: [SPM] In need of DTI help
>>
>> Dear Jessica,
>>
>> do I understand you correctly that you have 3 different behavioural
>> scores
>> per subject and want to find whether there are areas in FA maps of a
>> group of subjects where anisotropy changes in relation to the
behaviour
>> scores? If so, yes this is technically possible with SPM - however
you
>> need to consider that FA is a quantity limited to a range between 0
and
>> 1.
>> Correlating this with arbitrary unlimited scores will not give
>> meaningful results, but you should be fine with e.g. percent-range
>> scores.
>>
>> Volkmar
>>
>> On Mon, 31 Jul 2006, <Jessica> <Galgano> wrote:
>>
>>> Hello SPM experts!
>>>
>>> I am in desperate need of some help. If you have 3 different
>>> cognitive scores that were
>>> measured from 3 different cognitive tasks, can you then correlate
>>> these scores with each pixel's FA value (DTI output) in the DTI
>>> brain map? Could this be done using only SPM? If so, can someone
>>> explain how?
>>>
>>> Thank you!
>>> Jessica, NY
>>>
>>> "Voxel-based correlational analysis. For each slice position,
>>> analysis of the DTI data produces mean diffusivity, FA, and T2-
>>> weighted images, which are identical in anatomic position. The
>>> first step in image processing involved stripping of the skull and
>>> dura using an automated algorithm22 (Brain Extraction Tool; Oxford
>>> Centre for Functional Magnetic Resonance Imaging of the
>>> Brain; http://www.fmrib.ox.ac.uk/fsl/bet/index.html). For each
>>> subject,
>>> the T2-weighted images were then fitted to a symmetric
>>> echo-planar MRI brain template using a 12-parameter, affine
>>> normalization
>>> algorithm from Statistical Parametric Mapping23 (SPM
>>> 99, Functional Imaging Laboratory, Institute of Neurology,
>>> University
>>> College London, UK; http//www.fil.ion.ucl.ac.uk/spm). A
>>> symmetric echo-planar template was used. An identical transformation
>>> was then applied to the mean diffusivity and FA images.
>>> This was repeated for all subjects so that a full set of DT images
>>> was produced, all fitted to an identical template. All of the
>>> normalized
>>> images were reviewed visually to ensure that there were
>>> no obvious registration errors. The normalized images were then
>>> smoothed using an isotropic Gaussian filter (full width half maximum
>>> 4 mm). This process reduces the impact of small errors in
>>> registration by recalculating the intensity at each voxel based on a
>>> weighted mean of intensity at that voxel and surrounding voxels."
>>>
>>>
>>
>
|
