help
help > RE: import condition/session specific values
Mar 9, 2016 04:03 AM | Alfonso Nieto-Castanon - Boston University
RE: import condition/session specific values
Dear Pauline,
Yes, if you have two conditions and look at the [-1 1] contrast you should get exactly the same results in CONN as you would get in SPM for the same within-subjects comparison (note: using either SPM's paired t-test equal variances analyses, or using SPM's two-stage approach of first computing subject-level contrast volumes and then entering those into a second-level one-sample t-test; both of these are exactly equivalent and CONN is using the latter approach). To clarify the degrees of freedom issue, if you have N subjects and you use SPM's paired t-test implementation, SPM concatenates the data of the two conditions and defines the paired t-test analysis by creating a design matrix with 2*N rows (the first N correspond to the first condition data for all subjects, and the next N correspond to second condition data again for all subjects) and 2+N columns (the first two columns represent the two groups/conditions, and next N columns represent the subject-effects). This will produce a model with N-1 degrees of freedom (dof equals to the number of rows in the design matrix minus the rank of the design matrix, in this case "2*N - (N+1)"; note: the rank is N+1 instead of N+2, the number of columns, because the first two columns add up to the same as the last N columns in the design matrix). The analysis results, statistics, and degrees of freedom are exactly the same as those that you could get if you, instead, first compute subject-specific between-condition contrasts (i.e. define and estimate the "condition2" minus "condition1" within-subject effects directly on each subject separately), and then enter those N contrast volumes into a second-level one-sample t-test. In this latter case (and this is CONN's implementation), the second-level one-sample t-test analysis will be simply using a design matrix with N rows and a single column (containing 1's for every subject), and the degrees of freedom of these analyses would also be N-1 (and in fact not only the degrees of freedom will be the same in the two analyses -SPM's paired t-test implementation and SPM/CONN's two-stage implementation-, all the T-statistics, effect sizes, and p-values will also be exactly the same, as those two analyses simply represent two equivalent representations of the same model and hypothesis being tested).
Hope this helps
Alfonso
Originally posted by Pauline Favre:
Yes, if you have two conditions and look at the [-1 1] contrast you should get exactly the same results in CONN as you would get in SPM for the same within-subjects comparison (note: using either SPM's paired t-test equal variances analyses, or using SPM's two-stage approach of first computing subject-level contrast volumes and then entering those into a second-level one-sample t-test; both of these are exactly equivalent and CONN is using the latter approach). To clarify the degrees of freedom issue, if you have N subjects and you use SPM's paired t-test implementation, SPM concatenates the data of the two conditions and defines the paired t-test analysis by creating a design matrix with 2*N rows (the first N correspond to the first condition data for all subjects, and the next N correspond to second condition data again for all subjects) and 2+N columns (the first two columns represent the two groups/conditions, and next N columns represent the subject-effects). This will produce a model with N-1 degrees of freedom (dof equals to the number of rows in the design matrix minus the rank of the design matrix, in this case "2*N - (N+1)"; note: the rank is N+1 instead of N+2, the number of columns, because the first two columns add up to the same as the last N columns in the design matrix). The analysis results, statistics, and degrees of freedom are exactly the same as those that you could get if you, instead, first compute subject-specific between-condition contrasts (i.e. define and estimate the "condition2" minus "condition1" within-subject effects directly on each subject separately), and then enter those N contrast volumes into a second-level one-sample t-test. In this latter case (and this is CONN's implementation), the second-level one-sample t-test analysis will be simply using a design matrix with N rows and a single column (containing 1's for every subject), and the degrees of freedom of these analyses would also be N-1 (and in fact not only the degrees of freedom will be the same in the two analyses -SPM's paired t-test implementation and SPM/CONN's two-stage implementation-, all the T-statistics, effect sizes, and p-values will also be exactly the same, as those two analyses simply represent two equivalent representations of the same model and hypothesis being tested).
Hope this helps
Alfonso
Originally posted by Pauline Favre:
Thanks Alfonso, that helps a lot!
May I ask another question? I was also wondering if the way that within-subject analysis as performed in conn is the same as paired-t-tests in spm?
because in the design matrix I got only one column for the pre vs post effect and one CON.nii file per subject (are those correspond to the residual error or the within-subjects effect?), also the results should not be exactly the same as in spm since the number of dof is different? (i.e., only one regressor in conn vs two or one per condition + one per subject in spm)?
Many thanks
Best
Pauline
May I ask another question? I was also wondering if the way that within-subject analysis as performed in conn is the same as paired-t-tests in spm?
because in the design matrix I got only one column for the pre vs post effect and one CON.nii file per subject (are those correspond to the residual error or the within-subjects effect?), also the results should not be exactly the same as in spm since the number of dof is different? (i.e., only one regressor in conn vs two or one per condition + one per subject in spm)?
Many thanks
Best
Pauline
Threaded View
| Title | Author | Date |
|---|---|---|
| Pauline Favre | Mar 8, 2016 | |
| Alfonso Nieto-Castanon | Mar 8, 2016 | |
| Pauline Favre | Mar 8, 2016 | |
| Alfonso Nieto-Castanon | Mar 9, 2016 | |