open-discussion
open-discussion > RE: Resting-state seed connectivity analysis
Sep 11, 2015 05:09 PM | Alfonso Nieto-Castanon - Boston University
RE: Resting-state seed connectivity analysis
Hi,
Just to add to the discussion, if you are interested in bringing these connectivity measures to a second-level analysis across subjects, then:
a) if you are going to be entering your regression coefficient measures (beta* maps) to your second-level model, then both approaches will be exactly equivalent
b) if you are going to be entering correlation coefficient measures (e.g. computing Fisher-transformed r) to your second-level model, then the first approach (your original suggestion) is the correct/simpler way to go (e.g. you can compute correlation coefficients from the regression coefficients simply by multiplying the regression coefficients by the ratio of standard deviations in your seed and target regions; if you use the latter approach this same computation will give you part/semi-partial correlation measures, instead of the more widely used partial correlation measures)
Cheers
Alfonso
ps. if you do not mind the biased recommendation (and you have access to Matlab), you may want to consider using the CONN toolbox (www.nitrc.org/projects/conn) for doing these analyses, that integrates nicely with SPM and it will handle all these issues automatically for you (e.g. removing motion and White/CSF confounds, scrubbing, filtering, etc.).
Originally posted by Jean-Baptiste Poline:
Just to add to the discussion, if you are interested in bringing these connectivity measures to a second-level analysis across subjects, then:
a) if you are going to be entering your regression coefficient measures (beta* maps) to your second-level model, then both approaches will be exactly equivalent
b) if you are going to be entering correlation coefficient measures (e.g. computing Fisher-transformed r) to your second-level model, then the first approach (your original suggestion) is the correct/simpler way to go (e.g. you can compute correlation coefficients from the regression coefficients simply by multiplying the regression coefficients by the ratio of standard deviations in your seed and target regions; if you use the latter approach this same computation will give you part/semi-partial correlation measures, instead of the more widely used partial correlation measures)
Cheers
Alfonso
ps. if you do not mind the biased recommendation (and you have access to Matlab), you may want to consider using the CONN toolbox (www.nitrc.org/projects/conn) for doing these analyses, that integrates nicely with SPM and it will handle all these issues automatically for you (e.g. removing motion and White/CSF confounds, scrubbing, filtering, etc.).
Originally posted by Jean-Baptiste Poline:
Hi,
If you are creating t-maps or estimated parameters maps the two approaches should be equivalent (if we ignore degrees of freedom ... ) since the t test will only test for the part of your seed region signal that is uncorrelated to the other parts of the model. The values that should change are the ones of the counfounds (WM CSF movement signals). I would therefore use the approach that you have seen in the litterature so that you have a better estimation of the variance of your parameters (correct df).
cheers
JB
If you are creating t-maps or estimated parameters maps the two approaches should be equivalent (if we ignore degrees of freedom ... ) since the t test will only test for the part of your seed region signal that is uncorrelated to the other parts of the model. The values that should change are the ones of the counfounds (WM CSF movement signals). I would therefore use the approach that you have seen in the litterature so that you have a better estimation of the variance of your parameters (correct df).
cheers
JB
Threaded View
| Title | Author | Date |
|---|---|---|
| Andrew Song | Aug 28, 2015 | |
| ladan shahshahani | Oct 30, 2015 | |
| Jean-Baptiste Poline | Sep 11, 2015 | |
| Alfonso Nieto-Castanon | Sep 11, 2015 | |