Dear Matthew,
Thank you very much for your help. Your explanation is very helpful for me in understanding some statistical concepts.
If I understand correctly, since the Hotelling-Williams test uses t and z statistic, it indeed assumes the parameters (for each ROI) to be normally-distributed and need bonferroni to correct for multiple comparisons as well (quite different from the permutation inference in GLM for imaging).
Best,
Ling
On Thu, 14 Nov 2019 11:44:54 +0000, Matthew Webster <[log in to unmask]> wrote:
>Dear Ling,
>I’m not familiar with the test mentioned, so I’ll leave that for others to answer more fully. But if the Hotelling-Williams test assumes normally-distributed errors, then that could be one source of difference.
>
>Your interaction contrasts look fine. The post-hoc results will likely be very similar, but not identical ( e.g. the extra data potentially allows for a different estimate of the null-distribution ).
>
>Kind Regards
>Matthew
>
>
>--------------------------------
>Dr Matthew Webster
>FMRIB Centre
>John Radcliffe Hospital
>University of Oxford
>
>On 13 Nov 2019, at 23:33, Ling <[log in to unmask]<mailto:[log in to unmask]>> wrote:
>
>Dear Matthew,
>
>Thank you very much for your help!
>
>I did what you suggested but got no significant results. However, if I use ROI approach and extract mean FA for some tracts and compared the associations between meanFA - A and meanFA - B using Hotelling-Williams test, it does show significant differences for some tracts.
>
>So I am a bit confused, whether the two methods, 1) to compare correlation coefficients in GLM as you suggested and 2) to compare them using Hotelling-Williams test (http://crr.ugent.be/archives/546), are statistically the same or not.
>
>My second question is, I have two groups (for example, group1 5 patients and group2 5 controls), and want to test the interaction between groups (patients or control) and my variables (A or B).
>
>According to your suggestion, I will apply this design matrix:
>Group = group (10 ones)
>EV1 = group 1 (5 ones 5 zeros)
>EV2 = group 2 (5 zeros 5 ones)
>EV3 = group 1 * variable A
>EV4 = group 2 * variable A
>EV5 = group 1 * variable B
>EV6 = group 2 * variable B
>EV7 = nuisance variable (age)
>
>my contrasts for testing the interaction between group (patients or control) and my variable (A or B) will be:
>two t tests (p=0.025)
>0 0 0 1 -1 -1 1 0
>0 0 0 -1 1 1 -1 0
>
>and the post hoc tests will be
>the difference between FA - A and FA - B for group1 (p=0.025)
>0 0 0 1 0 -1 0 0
>0 0 0 -1 0 1 0 0
>
>the difference between FA - A and FA - B for group2 (p=0.025)
>0 0 0 0 1 0 -1 0
>0 0 0 0 -1 0 1 0
>
>If I use GLM, are these design matrix and contrasts correct? The results that I will get from the two post-hoc tests would be the same if I run TBSS for two groups separately (two 4D files for each groups, if we don't consider the different skeletons of two groups), am I right?
>
>I would appreciate any of your help.
>
>Ling
>
>
>On Wed, 13 Nov 2019 13:51:45 +0000, Matthew Webster <[log in to unmask]<mailto:[log in to unmask]>> wrote:
>
>Dear Ling,
>Unless there is a strong reason not to, you can just include A and B in the same model and run a 1,-1 contrast to look for significant differences between them.
>
>Hope this helps,
>Kind Regards,
>Matthew
>--------------------------------
>Dr Matthew Webster
>FMRIB Centre
>John Radcliffe Hospital
>University of Oxford
>
>On 12 Nov 2019, at 18:00, Ling <[log in to unmask]<mailto:[log in to unmask]><mailto:[log in to unmask]>> wrote:
>
>Dear experts,
>
>I ran two separate TBSS to correlate FA with two variables A and B (GLM model), and I want to compare whether these two correlation coefficients differ at a voxel-wise level.
>
>I was thinking to convert the two t-stat maps that I got from the two separated TBSS (FA-A and FA-B) into two r-stat maps and using Hotelling-Williams test (http://crr.ugent.be/archives/546) to get a single t-stat map, which represents the difference of the two associations at each voxel.
>
>Then the question is how to correct for multiple comparisons using permutation test? Since in order to run a permutation test, you need data points for each subjects. But here, for each voxel, there is only one value (a t value) .
>
>I wonder whether this way of comparing correlations for the same sample makes sense and how to proceed?
>
>Thank you very much for your help!
>
>Ling
>
>########################################################################
>
>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
|