help
help > RE: reporting 3-way interaction
Sep 2, 2016 08:09 AM | Helene Veenstra
RE: reporting 3-way interaction
This helps me very much, thank
you!
Originally posted by Alfonso Nieto-Castanon:
Originally posted by Alfonso Nieto-Castanon:
Dear
Helene,
Yes, that looks perfectly fine. I probably would have used the simpler (but equivalent)
Group: [-1 1] (Group1, Group2)
Conditions: [3 1 -1 -3 -3 -1 1 3] (preWM1, preWM2, preWM3, preWM4, postWM1, postWM2, postWM3, postWM4)
and then simply look at the two-sided results.
This would typically be referred to as a "Group by Intervention by WM" 3-way interaction. In other words, you are looking at the changes in connectivity associated with WM difficulty, evaluating whether those associations differ between pre- and post- intervention and testing whether those pre-post changes are different in your two groups.
To evaluate the simple main effects, for example your hypothesis that "there is a between-group difference in connectivity strength related to WM difficulty" (assumming here that you are testing this pre- intervention only) you would use instead something like:
Group: [ -1 1] (Group1, Group2)
Conditions: [-3 -1 1 3] (preWM1, ... preWM4)
For any of these tests would typically want to then import and plot the individual-condition data for any significant connection in order to help you interpret the found effects.
Hope this helps
Alfonso
Originally posted by Helene Veenstra:
Yes, that looks perfectly fine. I probably would have used the simpler (but equivalent)
Group: [-1 1] (Group1, Group2)
Conditions: [3 1 -1 -3 -3 -1 1 3] (preWM1, preWM2, preWM3, preWM4, postWM1, postWM2, postWM3, postWM4)
and then simply look at the two-sided results.
This would typically be referred to as a "Group by Intervention by WM" 3-way interaction. In other words, you are looking at the changes in connectivity associated with WM difficulty, evaluating whether those associations differ between pre- and post- intervention and testing whether those pre-post changes are different in your two groups.
To evaluate the simple main effects, for example your hypothesis that "there is a between-group difference in connectivity strength related to WM difficulty" (assumming here that you are testing this pre- intervention only) you would use instead something like:
Group: [ -1 1] (Group1, Group2)
Conditions: [-3 -1 1 3] (preWM1, ... preWM4)
For any of these tests would typically want to then import and plot the individual-condition data for any significant connection in order to help you interpret the found effects.
Hope this helps
Alfonso
Originally posted by Helene Veenstra:
Dear Alfonso and
others,
Having read and applied tips from below conversation I wanted to check whether my ANOVA is done correctly as well, as it is a bit more complicated and confusing.
I have two groups, two working memory conditions (before and after intervention) with 4 levels, and three sets of ROIs I wish to test for connectivity changes. Since I do not hypothesise that the ROIs differ from each other regarding connectivity strength this is rather a 2-way ANOVA if I'm not mistaken?
So the working memory task (WM) has four increasing levels of difficulty, which I want to model as [-3 -1 1 3].
If my hypothesis is that there is a between-group difference in connectivity strength related to WM difficulty; then I want to model an F-test as follows:
Group: GroupA, GroupB
Conditions: WM_beforeIntervention level 1:4 [-3 1 1 3], WM_afterIntervention level 1:4 [-3 1 1 3]
ROI: 3 sets of ROIs that I want to test the connectivity between
F-test for Difference between groups - F-test for difference between before/after condition - test ROI connectivity
Group [1 -1;-1 1] - Condition [-3 -1 1 3 3 1 -1 -3; 3 1 -1 -3 -3 -1 1 3] - ROI [1] (select 'only test the 6 selected ROIs').
I hope I was clear in explaining my contrast. I wonder if above contrast is the best way to test this hypothesis?
thanks for your time!
Helene
Originally posted by Alfonso Nieto-Castanon:
Having read and applied tips from below conversation I wanted to check whether my ANOVA is done correctly as well, as it is a bit more complicated and confusing.
I have two groups, two working memory conditions (before and after intervention) with 4 levels, and three sets of ROIs I wish to test for connectivity changes. Since I do not hypothesise that the ROIs differ from each other regarding connectivity strength this is rather a 2-way ANOVA if I'm not mistaken?
So the working memory task (WM) has four increasing levels of difficulty, which I want to model as [-3 -1 1 3].
If my hypothesis is that there is a between-group difference in connectivity strength related to WM difficulty; then I want to model an F-test as follows:
Group: GroupA, GroupB
Conditions: WM_beforeIntervention level 1:4 [-3 1 1 3], WM_afterIntervention level 1:4 [-3 1 1 3]
ROI: 3 sets of ROIs that I want to test the connectivity between
F-test for Difference between groups - F-test for difference between before/after condition - test ROI connectivity
Group [1 -1;-1 1] - Condition [-3 -1 1 3 3 1 -1 -3; 3 1 -1 -3 -3 -1 1 3] - ROI [1] (select 'only test the 6 selected ROIs').
I hope I was clear in explaining my contrast. I wonder if above contrast is the best way to test this hypothesis?
thanks for your time!
Helene
Originally posted by Alfonso Nieto-Castanon:
Dear
Bruno
If you want to use F-stats (and that is typically the best choice for this sort of interaction analyses, unless you have an a priori hypothesis about the directionality of the expected interaction effect) simply switch the option in the results explorer window that reads 'one-sided (positive)' to 'two-sided', and that will be exactly equivalent to the standard F-test (which is blind to the directionality of the interaction effect) for this same interaction analysis.
If in doubt, or if you prefer to have F-stat values directly, you could also simply enter in the 'between-subjects contrast' field [1 -1; 1 -1] (instead of just [1 -1]), and that will give you exactly the same results as the analysis above, but now all statistics will be reported using F-stats instead of T-stats. Another alternative would be, in the original analysis results explorer window, click on the 'display SPM' button and define there a new F-contrast with the [-1 1] values. Again, all of these options will produce exactly the same second-level results as your original analyses (using the 'two-sided' option), you will see exactly the same significant clusters, the same cluster-level statistics, etc. (they will only vary in the choice of voxel-level statistics being reported -either T(dof) or F(1,dof) stats-)
Hope this helps
Alfonso
Originally posted by Bruno Baumann:
If you want to use F-stats (and that is typically the best choice for this sort of interaction analyses, unless you have an a priori hypothesis about the directionality of the expected interaction effect) simply switch the option in the results explorer window that reads 'one-sided (positive)' to 'two-sided', and that will be exactly equivalent to the standard F-test (which is blind to the directionality of the interaction effect) for this same interaction analysis.
If in doubt, or if you prefer to have F-stat values directly, you could also simply enter in the 'between-subjects contrast' field [1 -1; 1 -1] (instead of just [1 -1]), and that will give you exactly the same results as the analysis above, but now all statistics will be reported using F-stats instead of T-stats. Another alternative would be, in the original analysis results explorer window, click on the 'display SPM' button and define there a new F-contrast with the [-1 1] values. Again, all of these options will produce exactly the same second-level results as your original analyses (using the 'two-sided' option), you will see exactly the same significant clusters, the same cluster-level statistics, etc. (they will only vary in the choice of voxel-level statistics being reported -either T(dof) or F(1,dof) stats-)
Hope this helps
Alfonso
Originally posted by Bruno Baumann:
Dear Alfredo,
after reading several posts I'm still a bit puzzled how to report the results of a mixed-design ANOVA.
I set up a 2x2x2 model (group, condition, roi). The 3-way interaction gives me a T-value instead of F-values which would be expected for a rmANOVA for example.
group: [1 -1]
condition: [1 -1]
roi: [1 -1]
If I understand correctly this is founded in the way results are calculated (t-tests for within-subject-effects on 1st-level, subsequent results in t-tests on 2nd level (between-subjects-effects))
If I want to report F-values is it the correct way to calculate the F=t^2 as indicated in that post (https://www.nitrc.org/forum/message.php?...)?
I hope the specifications are sufficient.
Thanks in advance and best wishes,
Bruno
after reading several posts I'm still a bit puzzled how to report the results of a mixed-design ANOVA.
I set up a 2x2x2 model (group, condition, roi). The 3-way interaction gives me a T-value instead of F-values which would be expected for a rmANOVA for example.
group: [1 -1]
condition: [1 -1]
roi: [1 -1]
If I understand correctly this is founded in the way results are calculated (t-tests for within-subject-effects on 1st-level, subsequent results in t-tests on 2nd level (between-subjects-effects))
If I want to report F-values is it the correct way to calculate the F=t^2 as indicated in that post (https://www.nitrc.org/forum/message.php?...)?
I hope the specifications are sufficient.
Thanks in advance and best wishes,
Bruno
Threaded View
| Title | Author | Date |
|---|---|---|
| Bruno Baumann | May 3, 2016 | |
| Alfonso Nieto-Castanon | May 3, 2016 | |
| Helene Veenstra | Aug 31, 2016 | |
| Alfonso Nieto-Castanon | Sep 1, 2016 | |
| Helene Veenstra | Sep 2, 2016 | |