help
help > RE: Multivariate regression - assumptions?
Oct 14, 2015 05:10 PM | Alfonso Nieto-Castanon - Boston University
RE: Multivariate regression - assumptions?
Hi Jenna,
Some thoughts on your questions below
Best
Alfonso
Originally posted by Jenna Traynor:
Yes, exactly.
With regard to the way I set it up, yes I did employ the steps that you suggested by using dummy-coding and making C1-C4 variables all 0 for control subjects.
However after doing step c) in the second-level results tab select 'ASD', 'C1' ... 'C4' in the subject-effects list and enter a between-subjects contrast [0 1 0 0 0; 0 0 1 0 0; 0 0 0 1 0; 0 0 0 0 1]
... would I scroll through the source ROIs on the right separately to see how my predictor variables are associated with each ROI? Or together, since the results are only giving me connectivity associated with ANY of the four C1 - C4 variables?
Sorry I missed the part in your original question where you had mentioned the multiple seed ROIs. If you want to look at each seed separatey then you would simply select each seed ROI individually in the 'sources' list, while if you want to look at the effects across any of your seeds (appropriately controlling for the number of seed ROIs) then you would select all of them simultaneously and enter an eye(N) contrast (the default contrast when selecting multiple seeds). If the number of seed ROIs is relatively large compared to your sample size then these analyses might not have sufficient degrees of freedom or sufficient power. In that case you may click on the 'results explorer' button, select there your multiple seed ROIs in the 'sources' list and use one of the additional thresholding options available there (e.g. seed-level FDR-corrected stats, network based statistics, etc.)
And for future - if I did ever want to employ multivariate regression would I also set up the second-level results in the same way? ie [0,1,0,0,0; 0,0,1,0,0; 0,0,0,1,0; 0,0,0,0,1]
Yes, exactly. In general, the way to define the second-level analyses is not affected by the choice of first-level connectivity measure entered in these analyses (e.g. bivariate/multivariate regression/correlation seed-to-voxel or ROI-to-ROI measures, but also PPI, voxel-to-voxel, graph measures, etc. after estimating any of those connectivity measures for each subject in the first-level analysis step they can then be analyzed across subjects using the same second-level general linear model definitions/contrasts)
Hope this helps
Alfonso
Some thoughts on your questions below
Best
Alfonso
Originally posted by Jenna Traynor:
Hi Alfonso,
Thank you so much for this clarification. Would it then be accurate to say that using multivariate regression is better when looking for the unique contribution of each ROI that is within a known network? For example, if I was looking at the DMN and wanted to see the unique contribution of the PCC after controlling for the contribution of the MPFC, LLP, etc, then it would be better to employ multivariate methods? But if examining a number of different a priori ROIs that are not necessarily a part of a FC network (as in my study; ie., insula, hippocampus, putamen) then it would be better to use separate regression models for each ROI (i.e., bivariate)?
Thank you so much for this clarification. Would it then be accurate to say that using multivariate regression is better when looking for the unique contribution of each ROI that is within a known network? For example, if I was looking at the DMN and wanted to see the unique contribution of the PCC after controlling for the contribution of the MPFC, LLP, etc, then it would be better to employ multivariate methods? But if examining a number of different a priori ROIs that are not necessarily a part of a FC network (as in my study; ie., insula, hippocampus, putamen) then it would be better to use separate regression models for each ROI (i.e., bivariate)?
Yes, exactly.
With regard to the way I set it up, yes I did employ the steps that you suggested by using dummy-coding and making C1-C4 variables all 0 for control subjects.
However after doing step c) in the second-level results tab select 'ASD', 'C1' ... 'C4' in the subject-effects list and enter a between-subjects contrast [0 1 0 0 0; 0 0 1 0 0; 0 0 0 1 0; 0 0 0 0 1]
... would I scroll through the source ROIs on the right separately to see how my predictor variables are associated with each ROI? Or together, since the results are only giving me connectivity associated with ANY of the four C1 - C4 variables?
Sorry I missed the part in your original question where you had mentioned the multiple seed ROIs. If you want to look at each seed separatey then you would simply select each seed ROI individually in the 'sources' list, while if you want to look at the effects across any of your seeds (appropriately controlling for the number of seed ROIs) then you would select all of them simultaneously and enter an eye(N) contrast (the default contrast when selecting multiple seeds). If the number of seed ROIs is relatively large compared to your sample size then these analyses might not have sufficient degrees of freedom or sufficient power. In that case you may click on the 'results explorer' button, select there your multiple seed ROIs in the 'sources' list and use one of the additional thresholding options available there (e.g. seed-level FDR-corrected stats, network based statistics, etc.)
And for future - if I did ever want to employ multivariate regression would I also set up the second-level results in the same way? ie [0,1,0,0,0; 0,0,1,0,0; 0,0,0,1,0; 0,0,0,0,1]
Yes, exactly. In general, the way to define the second-level analyses is not affected by the choice of first-level connectivity measure entered in these analyses (e.g. bivariate/multivariate regression/correlation seed-to-voxel or ROI-to-ROI measures, but also PPI, voxel-to-voxel, graph measures, etc. after estimating any of those connectivity measures for each subject in the first-level analysis step they can then be analyzed across subjects using the same second-level general linear model definitions/contrasts)
Hope this helps
Alfonso
Threaded View
| Title | Author | Date |
|---|---|---|
| Jenna Traynor | Oct 11, 2015 | |
| Alfonso Nieto-Castanon | Oct 11, 2015 | |
| Jenna Traynor | Oct 12, 2015 | |
| Alfonso Nieto-Castanon | Oct 14, 2015 | |
| Shruti Kinger | 22 hours ago | |
| Jenna Traynor | Oct 14, 2015 | |
| Alfonso Nieto-Castanon | Oct 14, 2015 | |
| Jenna Traynor | Oct 26, 2015 | |
| Jenna Traynor | Oct 28, 2015 | |