Diederick and John,


1st covariates: GM vs ICV ; you should check if results hold whenthere is no cov., for GM and for ICV as you pointed out already - if you want to keep GM and ICV into the analyses, orthogonalize one onto the other and enter them both.


2nd partial volumes - question for John ?? - given that the thalamus is supposed to be mainly GM - wouldn't that make sense to run the analysis again using a higher threshold ie keeping voxels with higher GM probability ? if that makes sense, then a plot of cluster size as a function of the GM thresholds (like 0.2 0. 4 0.6 0.8) would be a nice illustration, cluster size should be more or less the same despite being more conservative in term of GM prob. (or a normalized version like cluster size / total number of voxel in the pulvinar ROI)


Cyril





From: SPM (Statistical Parametric Mapping) <[log in to unmask]> on behalf of John Ashburner <[log in to unmask]>
Sent: 11 August 2015 13:03
To: [log in to unmask]
Subject: Re: [SPM] VBM criticism by reviewer
 
Partial volume between CSF and WM is often mis-classified as GM (because it has roughly the same intensities).  Larger ventricles have more surface area, so there could potentially be more of these partial volume voxels.  If one population has bigger ventricles than the other, you may see what appears to be more GM in regions close to the ventricles.

The statistical tests try to reject the null hypothesis that there's no difference between the preprocessed data. If you see a statistically significant difference, then (providing the design is reasonably balanced so that the statistical statistical tests are valid - see eg Salmond et al) you should be seeing a real difference between the preprocessed images.  However, there is always the possibility that these do not actually relate to GM ("correctly rejecting the null hypothesis for the wrong reason" - https://en.wikipedia.org/wiki/Type_III_error).

I guess you could do a CSF VBM to see if ventricles are actually bigger in one group.

Best regards,
-John


On 10 August 2015 at 22:36, Diederick Stoffers <[log in to unmask]> wrote:
Dear all,

In a VBM study we find a strong association between a clinical measure and pulvinar volume. We are able to replicate this in a second completely independent sample.

A reviewer is concerned about the location of this association.  He states:

The area with the greatest change in volume is adjacent to the atrium of the lateral ventricle, precisely where it begins to curve downward.  In other words, the surface of the pulvinar is surrounded by cerebrospinal fluid on three sides (dorsally, laterally, and caudal).  Structures along the ventricular surface are particularly liable to artifactual "loss of volume" when there are minute changes in ventricular volume, and this is in proportion to the degree to which they share surface with the ventricle.  Even a tiny change in overall brain volume is likely to result in the hot spot for loss of substance at the points where the brain tissue is surrounded by CSF.

We are correcting for total native GM in our analyses, not total brain volume. Could anyone comment on the validity of the reasoning of the reviewer? Would the way to go be to show results hold when controlling for total ICV or not controlling for brain volume at all?

Best,

Diederick