Dear AFNI experts,
My goal is to create whole-brain statistical maps for the interaction of brain data between two groups of individuals who received a different treatment and were scanned pre- and post-treatment. I’d like to see how the brain changed differentially between treatments.
I am not an expert in this aspect of AFNI and I see that there appear to be many options to do this, including 3dANOVA2, 3dMEMA, 3dLME, and perhaps others.
What would you suggest?
Dreaming of AFNI,
Matthew
Matthew,
So you have a 2 x 2 design with a between-subjects factor (two groups) and a within-subject factor (two conditions). Questions:
Dear Gang,
Thank you for your quick response! Here are answers to your questions:
1) Any missing data?
No missing data
2) Equal number of subjects across the two groups?
The groups are unequal in size, something like 18 and 14
3) Do you have t-statistic available for each effect estimate (beta)?
I was planning to run the model on z-score maps that were converted from pearson correlation maps. That is, I am interested in seeing how “resting-state functional connectivity” changes differentially based on treatment.
I am thinking to run this command:
3dMVM -prefix group_treatment_interaction -jobs 1 -bsVars Group -wsVars Time -dataTable @table.txt
Where “table.txt” is organized like this:
Subj Group Time InputFile
1 treatment_1 preTreatment /path_to_data/pretreatment_data.nii.gz
1 treatment_1 postTreatment /path_to_data/posttreatment_data.nii.gz
2 treatment_2 preTreatment /path_to_data/pretreatment_data.nii.gz
2 treatment_2 postTreatment /path_to_data/posttreatment_data.nii.gz
…
My undertanding is that 3dMVM computes interactions automatically between within and between subject effects, so I imagine an F-map output from this command will be what I am looking for. Correct?
Thank you for your help!
Matthew
My undertanding is that 3dMVM computes interactions automatically between within and between subject effects,
so I imagine an F-map output from this command will be what I am looking for. Correct?
Yes, that’s right. Alternatively you can use 3dANOVA3 -type 5, which is much faster.
Dear Gang,
Thank you for your help, I will move forward with one of 3dMVM or 3dANOVA3 (type 5).
Also I have a related question, I am particularly interested in testing what is sometimes called a “spreading interaction”. That is, I’d like to specifically test for an interaction that captures a difference in posttreatment resting-state functional connectivity (rsfc) between groups when there is no difference in pretreatment rsfc between groups. I’ve pasted below some text from a relevant paper that does the this analysis that I am interested in.
My sense is that there is a way to do this using the GLT options with 3dMVM, but I do not know the syntax. I’d like to set up the contrast such that the contrast weights for increases in treatment_1 relative to treatment_2, i.e., contrast weights for [−1 (pretreatment, treatment_1), −1 (pretreatment, treatment_2), 3 (post-treatment, treatment_1), −1 (post-treatment, treatment_2]. Also, I am interested in testing both whether the treatment_2 group has EITHER an increase OR a decrease compared to the treatment_1 group. I am thinking to run two GLTs and then correct the alpha p-threshold by two. Is there a way to include the increase OR decrease in a single GLT and thus avoid the use of multiple comparisons correction? I imagine the GLT could be specified by something like the following, but I am certain that it is incorrect:
-gltLabel 1 spreading_interaction_increasesTreatment1 -gltCode 1 ‘Group : -1treatment_1 -1treatment_1 Time : 1pretreatment 3posttreatment’ \
Many thanks for your help,
Matthew
To test study predictions in the brain, we specified a time by group spreading interaction contrast that tested for baseline to postintervention increases in rsFC in the HEM program relative to no change in the HER program from baseline to postintervention using contrast weights: [−1 (pre, HEM), −1 (pre, HER), 3 (post, HEM), −1 (post, HER)]. This t contrast models the specific hypothesized differential group change from baseline to posttreatment. The strength of this approach (relative to testing for significant voxels using the more standard overall F contrast or just comparing the two groups at posttreatment only) is that it tests the specific prediction that the mindfulness meditation program increases rsFC from baseline to posttreatment compared with no change in the relaxation group (as opposed to other types of interaction patterns that might be significant with an F contrast analysis). Note that this approach compares the mindfulness group at posttreatment with the average of the other cells in this 2 × 2 design, testing the spreading interaction prediction (and not other interaction patterns, e.g., crossover interactions).
Matthew,
I’d like to specifically test for an interaction that captures a difference in posttreatment resting-state functional
connectivity (rsfc) between groups when there is no difference in pretreatment rsfc between groups.
I would perform the following 5 tests, and use test 1 combined with tests 2 and 3 (or test 1 combined with tests 4 and 5) to tease apart all the scenarios (test 1 is essentially the same as the interaction F-test from 3dMVM except that it carries the directionality information):
-gltLabel 1 interaction -gltCode 1 ‘Group : 1treatment_1 -1treatment_2 Time : 1pretreatment -1posttreatment’
-gltLabel 2 interaction_t1 -gltCode 2 ‘Group : 1treatment_1 Time : 1pretreatment -1posttreatment’
-gltLabel 3 interaction_t2 -gltCode 3 'Group : 1treatment_2 Time : 1pretreatment -1posttreatment’
-gltLabel 4 interaction_pre -gltCode 4 ‘Group : 1treatment_1 -1treatment_2 Time : 1pretreatment’
-gltLabel 5 interaction_post -gltCode 5 'Group : 1treatment_1 -1treatment_2 Time : 1posttreatment’ \
Without seeing any justification for the weight of 3 in the spreading interaction specification, I fail to understand the logic of the approach.
Dear Gang,
I have ran the GLT codes at you have suggested them, i.e.:
-gltLabel 1 interaction -gltCode 1 ‘Group : 1treatment_1 -1treatment_2 Time : 1pretreatment -1posttreatment’ \
-gltLabel 2 interaction_t1 -gltCode 2 ‘Group : 1treatment_1 Time : 1pretreatment -1posttreatment’ \
-gltLabel 3 interaction_t2 -gltCode 3 'Group : 1treatment_2 Time : 1pretreatment -1posttreatment’ \
-gltLabel 4 interaction_pre -gltCode 4 ‘Group : 1treatment_1 -1treatment_2 Time : 1pretreatment’ \
-gltLabel 5 interaction_post -gltCode 5 'Group : 1treatment_1 -1treatment_2 Time : 1posttreatment’ \
My goal with this analysis was to test study predictions in the brain by specifying a time by group “spreading interaction” contrast that tested for pretreatment to posttreatment increases in rsFC in the treatment_2 group relative to no change in the treatment_1 group from pretreatment to posttreatment using contrast weights. I ran the above GLTs and the interaction_t2 exhibited a statistically significant cluster of voxels, but when I extracted the average z-scores for each individuals from the clusters and plotted averages for the groups for both time points I see that there is still a pretreatment difference between the groups that is driving the interaction, and no posttreatment difference between groups. I’m sorry if I wasn’t clear in my previous question, but this makes me think that the GLTs coded above are not actually testing for a spreading interaction. Am I correct? if so, is there a way to use 3dMVM to test a “spreading interaction”. If it might be helpful, I’ve pasted below text from a prior paper that described a spreading interaction that was implemented.
Many thanks for your help,
Matthew
TEXT FROM PRIOR STUDY
To test study predictions in the brain, we specified a time by group spreading interaction contrast that tested for baseline to postintervention increases in rsFC in the HEM program relative to no change in the HER program from baseline to postintervention using contrast weights: [−1 (pre, HEM), −1 (pre, HER), 3 (post, HEM), −1 (post, HER)]. This t contrast models the specific hypothesized differential group change from baseline to posttreatment. The strength of this approach (relative to testing for significant voxels using the more standard overall F contrast or just comparing the two groups at posttreatment only) is that it tests the specific prediction that the mindfulness meditation program increases rsFC from baseline to posttreatment compared with no change in the relaxation group (as opposed to other types of interaction patterns that might be significant with an F contrast analysis). Note that this approach compares the mindfulness group at posttreatment with the average of the other cells in this 2 × 2 design, testing the spreading interaction prediction (and not other interaction patterns, e.g., crossover interactions).
Matthew,
Sorry I previously misread what you’re looking for! For
[−1 (pre, HEM), −1 (pre, HER), 3 (post, HEM), −1 (post, HER)]
you may have to abandon 3dMVM, and do this:
Hope this helps!
The National Institute of Mental Health (NIMH) is part of the National Institutes of Health (NIH), a component of the U.S. Department of Health and Human Services.
Freedom of Information Act |
HHS Vulnerability Disclosure
National Institute of Mental
Health | National Institutes of Health | U.S.
Department of Health and Human Services |