Hi Anderson,Thank you for this detailed answer.One further question concerning design matrix of this correlation study between imaging modalities including a voxel-wise regressor:1) do I have to include a column of 1s with no demeaningOR2) No column of 1s and demeaning of variables + design (with --demean option) ?Best,MatthieuHi Matthieu,Please see below:Hi Anderson,1) Thank you for this explanation. So, in order to test correlation between voxel-wise variables PALM will compute t-maps and corrected p-maps associated showing significant t-values, right ?Yes.2) Once I got t-maps, I will use fslmaths with the link to the formulae you gave to me in order to compute partial correlation coefficients maps. But once I got these should I only to consider partial correlation coefficients where p is significant ?The formula gives a direct mapping t -> r, and vice-versa. The p-value for r is the same as for t. It's up to you to do the conversion only for significant voxels, or for all of them, and then perhaps mask the non-significant out.3) What's the difference between Pearson correlation coefficients and partial correlation coefficients ?I don't think Pearson considered nuisance variables in 1895 but the partial correlations are a direct extension of it. There is no conceptual difference, only the partial takes into account the effect of nuisance, whereas the original Pearson's method doesn't. The core idea is still his, though (or Galton's...).All the best,AndersonThanks for your lights !Best,MatthieuHi Matthieu,Only if you want to. You can:a) Test either positive or negative.b) Test either positive or negative, and include the option -twotail.c) Test both positive and negative, and correct for the fact that you are looking into two contrasts with the option -corrcon.For just 2 contrasts that are one the reverse of the other, the (a) and (b) take half of the time compared to (c), but (c) is better in general that it works for any arbitrary set of contrasts, not just one the negative of the other, and don't require much manipulation to find out whether significant results are positive or negative.From the description of the experiment it seems there is no reason to suspect that results would be necessarily directional so I think I'd use (b) or (c), with a preference towards (c) as it will allow less work for you later to make figures (but more work for the computer, which is what it is for anyway...)Hope this helps!All the best,AndersonHi Anderson,Thank you this helps !To test the correlation should I to define two contrasts for positive and negative correlation ?Best,Matthieu
2018-09-10 12:56 GMT+00:00 Anderson M. Winkler <[log in to unmask]>:Hi Matthieu,In the -evperdat you need to specify which positions in the design X and Z occupy. Something as:palm -i Y.nii -evperdat X.nii 2 -evperdat 3 -d ttest.csv -t tcontrasts.csv [...other options...]The ttest.csv would contain various EVs, and EV2 and EV3 would be "dummy" regressors that would be replaced by the voxelwise EVs specified by the -evperdat options. EV1 could be an intercept for example, and there could be additional EVs (EV4, EV5, etc).The tcontrasts.csv would have size compatible with the number of EVs then.Hope this helps!All the best,AndersonHi Anderson,Thank you for your answer. So, in my case where I want to test correlation between Y and X regressing out Z should I use basically:palm-i Y-evperdat X-evperdat Z-d ttest.csv-t tcontrasts.csvwhere ttest.csv is a file with only one column of 1 as intercept and tcontrasts.csv is a file with both 1 and -1 contrasts ?Best,Matthieu
2018-09-10 12:21 GMT+00:00 Anderson M. Winkler <[log in to unmask]>:Hi Matthieu,
The option "-pearson" isn't compatible with "-evperdat", which means you need to drop "-pearson" and compute the usual t-stat. Then, once you have it, it can be converted to a Pearson correlation coefficient using fslmaths. The formula is under the heading "Partial correlation coefficient", at this link: https://brainder.org/2015/03/04/all-glm-formulas/The option "-person" will likely be removed from future versions as it creates a whole lot of confusion. In fact, the most accurate p-values are produced with the t and F-statistics. Remember that PALM was created as the testbed for many things I was working on during my DPhil. The -pearson wasn't really intended to be used in practice. It is not wrong but it lacks some nice properties that t and F have, and it can be used only in more restricted situations. Consider not using it (but it's ok to convert t to r using the formula from the link).All the best,AndersonDear Anderson,I am stucked in how defining PALM design/contrasts and command line options in order to compute Pearson's correlation maps between modality Y and modality X when regressing out a modality Z.Could you provide me help ?Best regards,Matthieu
To unsubscribe from the FSL list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=FSL&A=1
To unsubscribe from the FSL list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=FSL&A=1
To unsubscribe from the FSL list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=FSL&A=1
To unsubscribe from the FSL list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=FSL&A=1
To unsubscribe from the FSL list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=FSL&A=1
To unsubscribe from the FSL list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=FSL&A=1
To unsubscribe from the FSL list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=FSL&A=1
To unsubscribe from the FSL list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=FSL&A=1
To unsubscribe from the FSL list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=FSL&A=1
To unsubscribe from the FSL list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=FSL&A=1