difference between 3dDeconvolve results for individual regressor vs. GLT with 1 term

AFNI Message Board
1

I’m running the following command:
3dDeconvolve -input $subfolder/${f1}/medn.brainmask.blur6mm+tlrc $subfolder/${f2}/medn.brainmask.blur6mm+tlrc -polort A
-local_times
-num_stimts 11
-stim_times 1 /home/shared/aclab-fmri/Studies/33_MOTIP2018/MOTIP18_behav_data/INTERAM/${nicesub}/SA.1D ‘SPMG2’
-stim_label 1 SA
-stim_times 2 /home/shared/aclab-fmri/Studies/33_MOTIP2018/MOTIP18_behav_data/INTERAM/${nicesub}/WM0F.1D ‘SPMG2’
-stim_label 2 WM0F
-stim_times 3 /home/shared/aclab-fmri/Studies/33_MOTIP2018/MOTIP18_behav_data/INTERAM/${nicesub}/WM0M.1D ‘SPMG2’
-stim_label 3 WM0M
-stim_times 4 /home/shared/aclab-fmri/Studies/33_MOTIP2018/MOTIP18_behav_data/INTERAM/${nicesub}/WM1F0.1D ‘SPMG2’
-stim_label 4 WM1F0
-stim_times 5 /home/shared/aclab-fmri/Studies/33_MOTIP2018/MOTIP18_behav?data/INTERAM/${nicesub}/WM1F1.1D ‘SPMG2’
-stim_label 5 WM1F1
-stim_times 6 /home/shared/aclab-fmri/Studies/33_MOTIP2018/MOTIP18_behav_data/INTERAM/${nicesub}/WM1M0.1D ‘SPMG2’
-stim_label 6 WM1M0
-stim_times 7 /home/shared/aclab-fmri/Studies/33_MOTIP2018/MOTIP18_behav_data/INTERAM/${nicesub}/WM1M1.1D ‘SPMG2’
-stim_label 7 WM1M1
-stim_times 8 /home/shared/aclab-fmri/Studies/33_MOTIP2018/MOTIP18_behav_data/INTERAM/${nicesub}/WM2F0.1D ‘SPMG2’
-stim_label 8 WM2F0
-stim_times 9 /home/shared/aclab-fmri/Studies/33_MOTIP2018/MOTIP18_behav?data/INTERAM/${nicesub}/WM2F1.1D ‘SPMG2’
-stim_label 9 WM2F1
-stim_times 10 /home/shared/aclab-fmri/Studies/33_MOTIP2018/MOTIP18_behav_data/INTERAM/${nicesub}/WM2M0.1D ‘SPMG2’
-stim_label 10 WM2M0
-stim_times 11 /home/shared/aclab-fmri/Studies/33_MOTIP2018/MOTIP18_behav_data/INTERAM/${nicesub}/WM2M1.1D ‘SPMG2’
-stim_label 11 WM2M1
-gltsym ‘SYM: +WM0F +WM0M +WM1F0 +WM1F1 +WM1M0 +WM1M1 +WM2F0 +WM2F1 +WM2M0 +WM2M1 -SA’
-glt_label 1 wm_vs_sa
-gltsym ‘SYM: +WM2M1 +WM2M0 +WM2F1 +WM2F0 -WM0F -WM0M’
-glt_label 2 2back_vs_0back
-gltsym ‘SYM: +WM1M1 +WM1M0 +WM1F1 +WM1F0 -WM0F -WM0M’
-glt_label 3 1back_vs_0back
-gltsym ‘SYM: +SA’
-glt_label 4 SA
-gltsym ‘SYM: +WM0F +WM0M +WM1F0 +WM1F1 +WM1M0 +WM1M1 +WM2F0 +WM2F1 +WM2M0 +WM2M1’
-glt_label 5 WM
-tout -xjpeg ${nicesub}_20200218model -x1D ${nicesub}_20200218model
-bucket ${nicesub}_deconvolution_bucket \

When looking at the resulting bucket dataset, I find that the coefficients for the SA regressor (stim_label 1) are very different from GLT 4, which includes only the SA regressor. Why are they different? What is the GLT test doing in this case? I included it because the results of my other GLTs look really good, as expected, but the results for the SA regressor look weak and confusing - not what I expected - and I was confused, and testing things out to see what would happen.

2

Hello,

In this case, they will not be the same. Your SA term is modeled by 2 regressors, not just 1. In such a case, the unspecified GLT will be evaluated as the sum of the components, meaning ‘SYM: +SA’ is the same as ‘SYM: +SA[0] +SA[1]’. That sum is likely to wash out the individual components, to some degree.

If you want to compare directly, depending on which component you actually care about (likely the first), consider using something like:

-gltsym 'SYM: SA[0]' -glt_label 4 SA_0
  • rick
3

Thank you so much for explaining. I understand now why the “single variable” GLT was not actually single variable; it includes both SPMG2 parameters.

If I use


-gltsym 'SYM: SA[0]' -glt_label 4 SA_0

, why would that be different from just looking at the first parameter estimates for the SA regressor? What does doing the GLT add in this case if there is no contrast? I have certainly read that most people only use the first parameter, so I’m particularly concerned that my results for that parameter alone don’t look good, but the result for the two combined looks great.

4

Yes, that SA[0] GLT should be identical to the first parameter results the come by default. That was just to verify that the GLTs are not doing anything special.

To get a comparison that is more like the F-test for the SA terms, then ‘SYM: SA[0] \ SA[1]’ might be reasonable. That is a 2-line contrast for those two parameters (\ is used as a line separator), testing more whether either one is significant. But again, you probably already have an F-stat that corresponds with original term (or maybe not, since there does not seem to be any -fout in the 3dDeconvolve command).

Note that something like an F-test might not help so much here, assuming you will use single subject betas for a group analysis. Maybe it would be good to consider the 2-line contrast applied at the group level, instead.

  • rick
5

I meant to respond again to say thank you so much for your help and clarification!