help
help > RE: Clarification on contrasts in CONN 2nd-level multivariate analysis
Oct 5, 2018 10:10 AM | Alfonso Nieto-Castanon - Boston University
RE: Clarification on contrasts in CONN 2nd-level multivariate analysis
Dear Martyn,
That looks exactly correct. If you have one within-subject factor (pre- vs. post- scans; coded as two conditions in CONN), three components from a MVPA analyses, and two additional between-subject factors (coded as four covariates, each one encoding each single-cell in a 2x2 design), then if you look at the three factor interaction the second-level analysis will be exactly as you describe (in CONN you would select the four covariates in the 'subject effects' list and enter a [1 -1 -1 1] contrast; select the two conditions in the 'conditions' list and enter a [1 -1] contrast, and select the three MVPA components in the 'sources' list and leave the default [1 0 0;0 1 0;0 0 1] contrast there). In order to have CONN produce the exact same results as your own implementation please make sure to select the 'non-parametric' statistics options in the results explorer window (so that it uses the Wilks lambda statistics there instead of SPM's reml procedure). If that is still producing different results perhaps click on the "design" button in CONN's gui to look at your second-level analysis design and C/M contrasts to make sure everything looks identical to your own analyses. Note that CONN will have reduced the design so that it enters into the second-level model 3 maps directly containing the differences between pre- and post- in each of the three MVPA components and then uses a M=[1 0 0;0 1 0; 0 0 1] contrast matrix (instead of entering the 6 original maps and then entering a M=[1 -1 0 0 0 0;0 0 1 -1 0 0;0 0 0 0 1 -1] contrast matrix), but those two approaches will lead to exactly the same Wilks lambda statistics either way, so there must be some other difference between your two versions of the analyses.
Hope this helps
Alfonso
Originally posted by Martyn McFarquhar:
That looks exactly correct. If you have one within-subject factor (pre- vs. post- scans; coded as two conditions in CONN), three components from a MVPA analyses, and two additional between-subject factors (coded as four covariates, each one encoding each single-cell in a 2x2 design), then if you look at the three factor interaction the second-level analysis will be exactly as you describe (in CONN you would select the four covariates in the 'subject effects' list and enter a [1 -1 -1 1] contrast; select the two conditions in the 'conditions' list and enter a [1 -1] contrast, and select the three MVPA components in the 'sources' list and leave the default [1 0 0;0 1 0;0 0 1] contrast there). In order to have CONN produce the exact same results as your own implementation please make sure to select the 'non-parametric' statistics options in the results explorer window (so that it uses the Wilks lambda statistics there instead of SPM's reml procedure). If that is still producing different results perhaps click on the "design" button in CONN's gui to look at your second-level analysis design and C/M contrasts to make sure everything looks identical to your own analyses. Note that CONN will have reduced the design so that it enters into the second-level model 3 maps directly containing the differences between pre- and post- in each of the three MVPA components and then uses a M=[1 0 0;0 1 0; 0 0 1] contrast matrix (instead of entering the 6 original maps and then entering a M=[1 -1 0 0 0 0;0 0 1 -1 0 0;0 0 0 0 1 -1] contrast matrix), but those two approaches will lead to exactly the same Wilks lambda statistics either way, so there must be some other difference between your two versions of the analyses.
Hope this helps
Alfonso
Originally posted by Martyn McFarquhar:
Dear Alfonso,
I am currently working with some colleagues on a data analysis using CONN. They have sent me the individual subject component maps from an MVPA analysis, and have reported 2nd-level results using these components in CONN, but I have been unable to replicate their results using my own MANOVA implementation. I was just wondering whether you could clarify exactly what CONN is doing so that I can hopefully replicate this.
For context, the analysis is based on the 2 x 2 x 2 mixed ANOVA example on page 26 of the CONN manual. There are pre- and post-treatment scans and so each subject has a set of components for the pre and a set for the post. Based on this I was assuming that CONN was running some sort of omnibus test across components, so the contrasts would be something like (using 3 components as an example):
C = [1 -1 -1 1]
M = [1 -1 0 0 0 0
0 0 1 -1 0 0
0 0 0 0 1 -1]
such that you would get a multivariate test across components on the difference between pre and post, however, this does not seem to be producing the same results in my own implementation (using Wilks Lambda as the test statistic).
Could you please clarify how the omnibus test from across components is formulated in CONN and whether I have misunderstood anything about the CONN implementation.
Best wishes,
Martyn McFarquhar
I am currently working with some colleagues on a data analysis using CONN. They have sent me the individual subject component maps from an MVPA analysis, and have reported 2nd-level results using these components in CONN, but I have been unable to replicate their results using my own MANOVA implementation. I was just wondering whether you could clarify exactly what CONN is doing so that I can hopefully replicate this.
For context, the analysis is based on the 2 x 2 x 2 mixed ANOVA example on page 26 of the CONN manual. There are pre- and post-treatment scans and so each subject has a set of components for the pre and a set for the post. Based on this I was assuming that CONN was running some sort of omnibus test across components, so the contrasts would be something like (using 3 components as an example):
C = [1 -1 -1 1]
M = [1 -1 0 0 0 0
0 0 1 -1 0 0
0 0 0 0 1 -1]
such that you would get a multivariate test across components on the difference between pre and post, however, this does not seem to be producing the same results in my own implementation (using Wilks Lambda as the test statistic).
Could you please clarify how the omnibus test from across components is formulated in CONN and whether I have misunderstood anything about the CONN implementation.
Best wishes,
Martyn McFarquhar
Threaded View
| Title | Author | Date |
|---|---|---|
| Martyn McFarquhar | Oct 4, 2018 | |
| Alfonso Nieto-Castanon | Oct 5, 2018 | |
| Martyn McFarquhar | Oct 5, 2018 | |
| Alfonso Nieto-Castanon | Oct 5, 2018 | |
| Martyn McFarquhar | Oct 8, 2018 | |
| Alfonso Nieto-Castanon | Oct 8, 2018 | |
| Martyn McFarquhar | Oct 9, 2018 | |
| Alfonso Nieto-Castanon | Oct 9, 2018 | |
| Ali Amad | Oct 11, 2018 | |
| Alfonso Nieto-Castanon | Oct 12, 2018 | |
| Ali Amad | Oct 19, 2018 | |
| Martyn McFarquhar | Oct 11, 2018 | |
| Alfonso Nieto-Castanon | Oct 12, 2018 | |
| Martyn McFarquhar | Oct 12, 2018 | |
| Alfonso Nieto-Castanon | Oct 12, 2018 | |
| Martyn McFarquhar | Oct 15, 2018 | |
| Martyn McFarquhar | Oct 5, 2018 | |