help
help > RE: Beta Values
May 1, 2023 11:05 PM | Olivier Brown
RE: Beta Values
Hi again,
Thanks for the help. Apologies if this is a rudimentary question, but how exactly am I to combine the NIFTI images to extract the values themselves? I'm still struggling to find an exact coefficient for the model, so I need a bit more guidance as to how to do that for reporting the statistics. Even when changing the contrasts for the simple main effects, as you recommended, I don't get beta values but only t statistics comparing time 1 to time 2 and the F value.
-Olivier
Originally posted by Alfonso Nieto-Castanon:
Thanks for the help. Apologies if this is a rudimentary question, but how exactly am I to combine the NIFTI images to extract the values themselves? I'm still struggling to find an exact coefficient for the model, so I need a bit more guidance as to how to do that for reporting the statistics. Even when changing the contrasts for the simple main effects, as you recommended, I don't get beta values but only t statistics comparing time 1 to time 2 and the F value.
-Olivier
Originally posted by Alfonso Nieto-Castanon:
Hi
Olivier,
In that design the slopes of the BDI effect on connectivity within each group are encoded implicitly in the contrasts:
[0 0 0 0 0 1 0]. % slope for groupA (subjects with 0's in "group1" variable)
[0 0 0 0 0 1 1]. % slope for groupB (subjects with 1's in "group1" variable)
and the difference between those two slopes is reflected in the interaction contrast that you used:
[0 0 0 0 0 0 1]. % difference between groupB and groupA slopes
So for simple main effects you could just recompute the same interaction analysis but now changing the contrasts to the values above in order to look at the individual effects within each group (the second analysis that you used is not doing exactly that, and it is also missing a constant term so it should be producing a warning message in CONN alerting you to that issue)
Last, coming back to the question about the regression coefficients, those are stored in the NIFTI images named beta_####.img (beta_0001 is the image of first regressor coefficients in your model, beta_0002 the second, etc.). So, alternatively, you could manually recompute the simple main effects above simply combining the corresponding beta images (e.g. slope for groupA will match the values in beta_0006 and slope for groupB will match the values resulting from adding beta_0006 and beta_0007)
Hope this helps
Alfonso
Originally posted by Olivier Brown:
In that design the slopes of the BDI effect on connectivity within each group are encoded implicitly in the contrasts:
[0 0 0 0 0 1 0]. % slope for groupA (subjects with 0's in "group1" variable)
[0 0 0 0 0 1 1]. % slope for groupB (subjects with 1's in "group1" variable)
and the difference between those two slopes is reflected in the interaction contrast that you used:
[0 0 0 0 0 0 1]. % difference between groupB and groupA slopes
So for simple main effects you could just recompute the same interaction analysis but now changing the contrasts to the values above in order to look at the individual effects within each group (the second analysis that you used is not doing exactly that, and it is also missing a constant term so it should be producing a warning message in CONN alerting you to that issue)
Last, coming back to the question about the regression coefficients, those are stored in the NIFTI images named beta_####.img (beta_0001 is the image of first regressor coefficients in your model, beta_0002 the second, etc.). So, alternatively, you could manually recompute the simple main effects above simply combining the corresponding beta images (e.g. slope for groupA will match the values in beta_0006 and slope for groupB will match the values resulting from adding beta_0006 and beta_0007)
Hope this helps
Alfonso
Originally posted by Olivier Brown:
Hi there,
I'm hoping to find clear instructions on how to extract regression beta coefficients from my model, if possible. I'll explain my process:
I ran an interaction analysis as follows:
Seed-to-voxel
Between-subjects contrast:
AllSubjects; group1; age; sex; handedness; BDI (quantitative scale); group1*BDI
0; 0; 0; 0; 0; 0; 1
Conditions
time 1; time 2
-1 1
Seed
DMN.LP(R)
This should be a multiple regression with an interaction factor, whereby I'm looking for interactions between group and BDI when predicting changes in connectivity from time 1 to time 2. I found a significant interaction. I extracted the mask of the significant cluster as a step towards computing simple main effects. For simple main effects, I rerun the within-group analyses to find their regression slopes as such:
Seed-to-voxel
Between-subjects contrast:
group1; age; sex; handedness; BDI (quantitative scale)
0; 0; 0; 0; 1
Conditions
time 1; time 2
-1 1
Seed
DMN.LP(R)
This should be a multiple regression, whereby I'm looking for the effects of BDI when predicting changes in connectivity from time 1 to time 2 within group 1. In other words, one of the group slopes from the previous interaction. No significant clusters were found. I then apply the exported mask from the interaction to determine the simple main effects. Now when I go to the results explorer, I only find the t values of effects at time 1 vs. time 2. This comparison of means seems to fall in line with an ANCOVA - so was this not a regression? Anyway, I'm still left with my main question of this post:
Where/how can I find the beta coefficient for this analysis (with exported mask) to explain the magnitude of the slope for my regression?
Thanks!
I'm hoping to find clear instructions on how to extract regression beta coefficients from my model, if possible. I'll explain my process:
I ran an interaction analysis as follows:
Seed-to-voxel
Between-subjects contrast:
AllSubjects; group1; age; sex; handedness; BDI (quantitative scale); group1*BDI
0; 0; 0; 0; 0; 0; 1
Conditions
time 1; time 2
-1 1
Seed
DMN.LP(R)
This should be a multiple regression with an interaction factor, whereby I'm looking for interactions between group and BDI when predicting changes in connectivity from time 1 to time 2. I found a significant interaction. I extracted the mask of the significant cluster as a step towards computing simple main effects. For simple main effects, I rerun the within-group analyses to find their regression slopes as such:
Seed-to-voxel
Between-subjects contrast:
group1; age; sex; handedness; BDI (quantitative scale)
0; 0; 0; 0; 1
Conditions
time 1; time 2
-1 1
Seed
DMN.LP(R)
This should be a multiple regression, whereby I'm looking for the effects of BDI when predicting changes in connectivity from time 1 to time 2 within group 1. In other words, one of the group slopes from the previous interaction. No significant clusters were found. I then apply the exported mask from the interaction to determine the simple main effects. Now when I go to the results explorer, I only find the t values of effects at time 1 vs. time 2. This comparison of means seems to fall in line with an ANCOVA - so was this not a regression? Anyway, I'm still left with my main question of this post:
Where/how can I find the beta coefficient for this analysis (with exported mask) to explain the magnitude of the slope for my regression?
Thanks!
Threaded View
| Title | Author | Date |
|---|---|---|
| Olivier Brown | Apr 28, 2023 | |
| Alfonso Nieto-Castanon | May 1, 2023 | |
| Olivier Brown | May 1, 2023 | |
| Alfonso Nieto-Castanon | May 5, 2023 | |