# FreeSurfer contrast vector compared to 3dMVM glt

Hi,

I have a fairly simple question but I want to make sure I am thinking about glt contrast the right way for 3dMVM. Just for example, let’s say we have two groups with two covariates, age and weight. So the initial mvm set up would include ‘-qVars “Age,Weight”’ and ‘-bsVars Group’

I have some examples of Freesurfer contrast vectors that I want to equate to glt contrasts.

To look at group effect in FS while “regressing out” the covariates, the contrast vector is 1 -1 0 0 0 0. Would this be the same as a gltCode of “Group : 1Group1 -1Group2”?

And if I wanted to look at the group difference in Age controlling for weight, I would use a gltCode of “Group : 1Group1 -1Group2 Age : “ ?

And finally, to look the interaction of age and weight with group differences, would the glt be “Group : 1Group1 -1Group2 Age : Weight : “ ?

I’ve looked through tutorials and all of the message board posts and can’t quite find an answer. I apologize if I have missed something and this has been answered elsewhere.

Thanks

First, check out if centering is an issue you may need to address: https://afni.nimh.nih.gov/pub/dist/doc/htmldoc/statistics/center.html

To look at group effect in FS while “regressing out” the covariates, the contrast vector is 1 -1 0 0 0 0.
Would this be the same as a gltCode of “Group : 1Group1 -1Group2”?

Yes, that’s the specification when the two covariates are “controlled/fixed” (not regressing out) at some specific values.

And if I wanted to look at the group difference in Age controlling for weight, I would use a gltCode of “Group : 1Group1 -1Group2 Age : “ ?

Right.

And finally, to look the interaction of age and weight with group differences, would the glt be “Group : 1Group1 -1Group2 Age : Weight : “ ?

Not sure what you mean by the interaction. One possibility is to properly center the two variables first, and then create another variable in the model that is the product of the two (with a center of 0).

Thank you for your reply! I also disagree with the terminology (hence the quotes), just wanted to stay consistent with what was on the FreeSurfer tutorials.

I will definitely center the qVars, I usually do the pre-checks and centering before inclusion in the MVM script but will look into these options within MVM.

Yes, your description is exactly what I had in mind. I was not sure if 3dMVM handled this interaction term automatically when two qVars are specified.

I have a couple of additional questions if you have the time.

1. The data that I am using is not fMRI data so I don’t have first level statistics (betas and test statistics), I only have a single brick per subject with z-scores. Does the MVM framework rely on having first level analyses or is this type of image ok?

2. Is there a way to extract diagnostics from the model? (e.g., multicollinearity, homoscedasticity, etc.)

Thanks very much for your time

I only have a single brick per subject with z-scores. Does the MVM framework rely
on having first level analyses or is this type of image ok?

The program 3dMVM does not care what you feed, but the issue is the interpretation. What are the z-scores: Z-statistic or Fisher-transformed z-scores based on correlation values? The latter is fine, but the former would be problematic.

1. Is there a way to extract diagnostics from the model? (e.g., multicollinearity, homoscedasticity, etc.)

If there is exact collinearity, 3dMVM would not run and you would get 0 values as output. By homoscedasticity do you mean different variances across groups? There is no easy way to diagnose the issue at the whole brain level. The best hope is that the incorporation of covariates would mitigate the severity of the problem.

Interesting, thank you for your reply. I currently have a dataset of “activation” maps that were derived from a separate analysis for subjects in two groups (diseased and healthy) and I am interested in looking at the difference in these maps while controlling for covariates. The maps are raw values so it is simply the Z-stat, not the Fisher-transform. It seems after reading your response, the correct analysis would be with 3dttest++ instead of 3dMVM because there is no way to generate first-level statistics for a Fisher’s Z. Is that correct?

Thanks again for your help and I apologize for any naivety. I just want to make sure I am using the programs responsibly.

the correct analysis would be with 3dttest++ instead of 3dMVM because there is no way to generate
first-level statistics for a Fisher’s Z. Is that correct?

The issue here is not about a specific program or model. Rather, a model should be built to explain the data of a physical measurement such as BOLD response, body weight, reaction time, etc. It is not clear how your Z-stat was derived. In general, you don’t construct a model on statistical values: a statistical value is some dimensionless measure, not a physical (dimensional) measure.

Of course, it just sounded like perhaps you were saying that first level statistics were necessary for the multivariate modeling performed in the program. Sorry for the misunderstanding.

The data I am working with are derived from a neural network framework through which class activation maps were calculated (range 0-0.5 per voxel). Since the scale of this value is variable for each subject (i.e. a highly predictive voxel might have a class activation of 0.1 for one subject and 0.4 for another) I just calculated the z score within each subject’s data in an attempt to normalize values for group statistics. In retrospect, I should’ve shared this earlier to avoid much of the confusion so I apologize. So the metric being used here is likelihood that the intensity in a voxel contributes to the final decision made by the neural network.

I was under the impression that MVM needed an initial contrast condition for a subject over the time period of an fMRI experiment, since that is what it was designed to do. I have used it several times in the past in this way to look at how first level condition minus baseline maps (from 3dREMLfit) differ between groups but I have never used it without that initial test, which of course generates the betas and test statistics. This makes much more sense to me since, as you said, the statistic tells us something about the goodness of fit and contrast (ie hemodynamics) of both conditions to the HRF (or whatever basis function you use to generate your fMRI maps). However, I was having trouble wrapping my head around generalizing the framework to other applications where the variance and effect size of the contrast within each subject is not already known.

Hopefully that makes more sense now.

Again, thank you very much for your time. I admire your work, Gang, and definitely see you as a voice of statistical reason in a field that badly needs it.

Thanks

Thanks for the clarifications. Any feedback about the usability of the programs is appreciated.