help
help > RE: No BOLD variance in first level?
Apr 14, 2015 08:04 PM | Fred Uquillas
RE: No BOLD variance in first level?
Thank you so much Alfonso. This thoroughly clarified the concern I
had.
All the best,
Fred
Originally posted by Alfonso Nieto-Castanon:
All the best,
Fred
Originally posted by Alfonso Nieto-Castanon:
Hi Fred,
First, one potential source of confusion in these "percentage variance explained" plots is that, in older versions of CONN (in 2013 and some of the earlier 2014 releases) these plots used to represent only the "unique variance" explained by each factor (the total variance explained by that factor alone, after removing all that could be explained by the other factors), while in newer versions (later 2014 releases and beyond) these plots represent the "total variance" (unique and shared) explained by each factor (the total variance explained by that factor, even if some portion of that variance would also be explained by other factors). While the change was meant to make the interpretation of these plots more intuitive (otherwise, for example, adding a new confounding effect would lead to decreases of the percentage variance explained by the other confounding effects, which was confusing), one confusing side-effect of this change is that now one would tend to see a lot more "variance explained" in the plots generated by the new versions compared to the plots generated by the old versions on the same data.
That said, yes, depending on the amount of subject-movement and/or physiological noise in the data it is not uncommon to see a lot of "total" variance explained by motion and/or White matter and CSF confounds (often several white matter, CSF, and motion parameters can be highly correlated, so none of this explained variance would appear if displaying only the "unique" variance explained by each factor, i.e. when displaying the same data using the older CONN versions). In general, during the Denoising step, you mainly want to make sure that the voxel-to-voxel correlation histogram plot looks reasonably centered and narrow (even if the percentage-variance plots show a lot of explained variance). If the correlation histogram does not look reasonably centered/narrow, then the "percentage-variance" plots can help guide you in identifying the main sources of confounding effects in your data so that you can plan accordingly (e.g. subject-motion effects typically show spatial patterns depending on the main direction of motion, and they typically appear in all CSF, White matter, and motion regressors, so if you see something like this then you may try adding additional derivative terms to the motion regressors and/or using a more conservative ART threshold in order to better account for these effects)
Hope this helps
Alfonso
Originally posted by Fred Uquillas:
First, one potential source of confusion in these "percentage variance explained" plots is that, in older versions of CONN (in 2013 and some of the earlier 2014 releases) these plots used to represent only the "unique variance" explained by each factor (the total variance explained by that factor alone, after removing all that could be explained by the other factors), while in newer versions (later 2014 releases and beyond) these plots represent the "total variance" (unique and shared) explained by each factor (the total variance explained by that factor, even if some portion of that variance would also be explained by other factors). While the change was meant to make the interpretation of these plots more intuitive (otherwise, for example, adding a new confounding effect would lead to decreases of the percentage variance explained by the other confounding effects, which was confusing), one confusing side-effect of this change is that now one would tend to see a lot more "variance explained" in the plots generated by the new versions compared to the plots generated by the old versions on the same data.
That said, yes, depending on the amount of subject-movement and/or physiological noise in the data it is not uncommon to see a lot of "total" variance explained by motion and/or White matter and CSF confounds (often several white matter, CSF, and motion parameters can be highly correlated, so none of this explained variance would appear if displaying only the "unique" variance explained by each factor, i.e. when displaying the same data using the older CONN versions). In general, during the Denoising step, you mainly want to make sure that the voxel-to-voxel correlation histogram plot looks reasonably centered and narrow (even if the percentage-variance plots show a lot of explained variance). If the correlation histogram does not look reasonably centered/narrow, then the "percentage-variance" plots can help guide you in identifying the main sources of confounding effects in your data so that you can plan accordingly (e.g. subject-motion effects typically show spatial patterns depending on the main direction of motion, and they typically appear in all CSF, White matter, and motion regressors, so if you see something like this then you may try adding additional derivative terms to the motion regressors and/or using a more conservative ART threshold in order to better account for these effects)
Hope this helps
Alfonso
Originally posted by Fred Uquillas:
Hi
Alfonso,
If I may ask, I'm wondering what the opposite of this scenario would mean? As in, what if for CSF or white matter confounds one sees a heck of a lot of variance explained throughout the brain display?
All the best,
Fred
Originally posted by Alfonso Nieto-Castanon:
If I may ask, I'm wondering what the opposite of this scenario would mean? As in, what if for CSF or white matter confounds one sees a heck of a lot of variance explained throughout the brain display?
All the best,
Fred
Originally posted by Alfonso Nieto-Castanon:
Dear
Diana,
Your snapshot of the preprocessing step looks perfectly fine, the lack of suprathreshold voxels in the display only indicates that the contribution of your confounding effects is below 50% of the total variance in the BOLD signal (this is not uncommon; you may change the threshold value located below the display to see the voxels where the percentage variance explained is above the specified threshold). Regarding the first-level connectivity maps, if you could attach a similar display I might be able to tell you with more confidence, but this could also (potentially) be just a matter of the interpretation of the display figure (try increasing the threshold value, which represents there uncorrected voxel-level p-values, or shifting through the image using the slider located at the right of the display, to see if you can see suprathreshold voxels for your different subjects/seeds). Let me know if this helps clarify a bit.
Best
Alfonso
Originally posted by Diana Wotruba:
Your snapshot of the preprocessing step looks perfectly fine, the lack of suprathreshold voxels in the display only indicates that the contribution of your confounding effects is below 50% of the total variance in the BOLD signal (this is not uncommon; you may change the threshold value located below the display to see the voxels where the percentage variance explained is above the specified threshold). Regarding the first-level connectivity maps, if you could attach a similar display I might be able to tell you with more confidence, but this could also (potentially) be just a matter of the interpretation of the display figure (try increasing the threshold value, which represents there uncorrected voxel-level p-values, or shifting through the image using the slider located at the right of the display, to see if you can see suprathreshold voxels for your different subjects/seeds). Let me know if this helps clarify a bit.
Best
Alfonso
Originally posted by Diana Wotruba:
Dear Conn users
I am having some trouble with running the first level analysis in the newest version of Conn (13f).
I loaded the realigned, slicetimed and normalized functional images (analyze format) and the corresponding coregistered and normalized segmented images (wc1*, wc2*, wc*3). This went perfectly fine for some but not all subjects. In appr. half of the subjects there seems to be no BOLD variance explained for the confounds (which is improbable). Still the vx2vx connectivity distribution looks fine (the original distribution shifted to the right, centered without the confounds).
(Also in the first level analysis results, there are no connectivity maps displayed for the same subjects)
(motion derivative order 1, dimension 6; white matter /CSF derivative order 0 dimension 3).
Any help would be very much appreciated!
Best
Diana
I am having some trouble with running the first level analysis in the newest version of Conn (13f).
I loaded the realigned, slicetimed and normalized functional images (analyze format) and the corresponding coregistered and normalized segmented images (wc1*, wc2*, wc*3). This went perfectly fine for some but not all subjects. In appr. half of the subjects there seems to be no BOLD variance explained for the confounds (which is improbable). Still the vx2vx connectivity distribution looks fine (the original distribution shifted to the right, centered without the confounds).
(Also in the first level analysis results, there are no connectivity maps displayed for the same subjects)
(motion derivative order 1, dimension 6; white matter /CSF derivative order 0 dimension 3).
Any help would be very much appreciated!
Best
Diana
Threaded View
| Title | Author | Date |
|---|---|---|
| Diana Wotruba | Mar 15, 2012 | |
| Kristina Gelardi | Mar 21, 2016 | |
| Alfonso Nieto-Castanon | Mar 31, 2016 | |
| Kristina Gelardi | Apr 6, 2016 | |
| Alfonso Nieto-Castanon | Mar 20, 2012 | |
| Fred Uquillas | Mar 26, 2015 | |
| Alfonso Nieto-Castanon | Apr 6, 2015 | |
| Fred Uquillas | Apr 14, 2015 | |
| Diana Wotruba | Mar 24, 2012 | |