help
help > RE: question about how denoising and 1st-level steps are implemented
Jun 26, 2018 05:06 PM | Alfonso Nieto-Castanon - Boston University
RE: question about how denoising and 1st-level steps are implemented
Hi Chaleece,
That is a good question, and yes, you are exactly right, the linear regression used as part of Denoising in CONN is a separate procedure from the linear regression used as part of seed-based correlation/regression first-level analyses. The main reasons why we are not doing this in a single regression step (e.g. like SPM does for functional activation analyses, where the effects of interest and the covariates of no interest are all combined in a single regression analysis) are:
a) because in functional connectivity analysis we generally need to remove those noise artifacts from both the "predictor" (e.g. seed BOLD timeseries) and "outcome" (e.g. voxel BOLD timeseries) variables in order to estimate proper (partial) correlation coefficient values (between the seed and target voxels), whereas in a typical functional analysis (e.g. SPM) we generally just need to remove those noise artifacts from the "outcome" variables (e.g. voxel BOLD timeseries) in order to estimate regression coefficient values (between the task effects and target voxels)
and b) because this allows us to use the same denoising procedure across multiple first-level connectivity estimation methods (including highly non-linear methods, where denoising might not be easily incorporated into the estimation procedure)
Hope this helps
Alfonso
Originally posted by Chaleece Sandberg:
That is a good question, and yes, you are exactly right, the linear regression used as part of Denoising in CONN is a separate procedure from the linear regression used as part of seed-based correlation/regression first-level analyses. The main reasons why we are not doing this in a single regression step (e.g. like SPM does for functional activation analyses, where the effects of interest and the covariates of no interest are all combined in a single regression analysis) are:
a) because in functional connectivity analysis we generally need to remove those noise artifacts from both the "predictor" (e.g. seed BOLD timeseries) and "outcome" (e.g. voxel BOLD timeseries) variables in order to estimate proper (partial) correlation coefficient values (between the seed and target voxels), whereas in a typical functional analysis (e.g. SPM) we generally just need to remove those noise artifacts from the "outcome" variables (e.g. voxel BOLD timeseries) in order to estimate regression coefficient values (between the task effects and target voxels)
and b) because this allows us to use the same denoising procedure across multiple first-level connectivity estimation methods (including highly non-linear methods, where denoising might not be easily incorporated into the estimation procedure)
Hope this helps
Alfonso
Originally posted by Chaleece Sandberg:
Hello good people!
I am posting a new question because I can't seem to find an answer after searching the forum (or the manual/tutorials).
Because denoising and the first-level analysis steps are implemented separately, I am assuming that the regression performed during the denoising step is separate from the GLM performed in the first-level analysis step. Is this an accurate assumption? I have been asked why everything is not simply combined into one GLM. Any explanation is greatly appreciated.
Thanks!
I am posting a new question because I can't seem to find an answer after searching the forum (or the manual/tutorials).
Because denoising and the first-level analysis steps are implemented separately, I am assuming that the regression performed during the denoising step is separate from the GLM performed in the first-level analysis step. Is this an accurate assumption? I have been asked why everything is not simply combined into one GLM. Any explanation is greatly appreciated.
Thanks!
Threaded View
| Title | Author | Date |
|---|---|---|
| Chaleece Sandberg | Jun 18, 2018 | |
| Alfonso Nieto-Castanon | Jun 26, 2018 | |
| Chaleece Sandberg | Jun 27, 2018 | |