AFNI version info (afni -ver):
Hello,
I have a question about the -AM2 command.
I used two components to add 3dDeconvolve, but there was a clear case where the beta became negative. We don't think that these ROIs will have negative activations when facing our tasks.
And, many times, one component's beta is positive and the other is negative. Is there anything I miss with the _AM2 command, or is it just the way my data is?
Thank you!
Hello,
With AM2 regression, there is a mean response component and differential components (where the mean has been removed). So keep in mind that even if the mean response is positive, the modulation terms need not be. A modulation term will be positive if an increase in the modulation value corresponds with an increase in the estimated BOLD response for that condition.
But also, a model can be set up with assumptions that do not quite fit the data. And a model with very many regressors can sometimes lead to surprising results, based on the interplay between sets of regressors.
A question of whether there is something odd about the results is often hard to answer, usually requiring scrutiny of both the model and data.
Thank you very much for your help! I understand the issue regarding positive and negative beta values now.
However, I still have a few questions. Currently, I am using two different components for AM2, meaning that the stimulus files are in the format x*a,b, where x is divided into seven groups corresponding to seven conditions.
I have two specific questions:
Thanks again for your help!
Comparing AM2 betas is much like comparing main betas, like scaling main regressors appropriately, the same must be done for modulators. Scaling will come from the magnitude of the modulators, so it is generally preferable for them to have a range that is in [-1,1], or something similar that might depend on the design. But you should scale them so that a 0.7 does mean the same magnitude of effect across modulators, for example, and after removing the mean. That is something you have to be careful about.
Of course there will also be the typical questions like correlations within the model. But scaling seems to be the main question here.
Dear rick,
Thanks for your help! But we still have some questions.
Our modulators for each group are in the same large range and have the same dimensions. However, the variances of the modulators for each group are not homogeneous, which makes us wonder whether the betas between groups are comparable.
In our understanding, AM2 is a regression based on the change of each group of modulators relative to the mean and the change of HRF. If the variances are not homogeneous, has AM2 adjusted the data? Can we compare betas? Or is there another way to deal with this problem?
Thanks again!
I don't actually see why the variance would be expected to match. You can only match so many things. If there are 2 stim classes where one class had twice the number of events, that regressors variance would surely be higher, yet people would still compare their betas. Of course since the scaling affects the variance, you could match variances that way, but the magnitude of the effect is what is generally important, not the variance directly. It is the magnitudes that would be compared with a group level test.
And of course it is important that the modulators are not too highly correlated.
Thank you for your help earlier! I’ve realized that matching the variances isn’t necessary for my analysis.
I now have another question. I have three different conditions and incorporated two factors, “a” and “b,” into the stimulus. Consequently, my stimulus file is structured as “2×3, 2.”
From my understanding, the #0 regressor represents the percent signal change across the different conditions. Meanwhile, the #1 regressors, Beta1 and Beta2, reflect the effects of factors “a” and “b” on changes in the hemodynamic response function.
I’m wondering: does the percent signal change represented by the #0 regressor remain independent of the effects of the two factors, “a” and “b”? In the other words, can I use this method to remove the effect of a and b to the percent signal change in one condition?
Unless you say otherwise, 3dDeconvolve will remove the mean from the modulators. One effect of that is that the modulators become basically orthogonal to the main term that they modulate. Note that including the a and b modulators does not necessarily even alter the results for the main component. That might depend on interactions with other regressors.
So regarding your last question, I would say yes, #0 is (mostly) independent of a and b, "mostly" because there are other regressors involved. Though your "in other words" statement is a little confusing.
-rick
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 |