Dear Alfonso

I have preprocessed 59 7T MRI resting-state data-sets for a ROI-to-ROI analyses. My interest goes to how two small nuclei are connected to other regions in the brain (using the AAL atlas). Therefore I had set the p-value at 0.025 FDR corrected (two-sided).
When using only the group factor (no covariates) I find many significant regions for the left and the right ROI of interest.
However, as soon as include a covariate, no matter which one  sex,education, ICV, QA_GCOR_rest) nothing survives the FDR threshold.
This is amazing, since many unadjusted effects were at p-FDR 0.000000 
To make sure this is not related to this specific small region, i did the same analyses for the precuneus (as this is an important hub) and I see similar effects: many significant ROI-to-ROI results before adjusting for covariates and nothing survives after adding covariates.

What might be happening here?

Two other questions:
- what is the QA_QCOR_rest representing?
- what is the difference between [0.5 0; 0 0.5] and [1 0; 0 1]?

many thanks!!
Best wishes
Heidi

Hello Heidi,

Just for the last question: the only difference it would have is that the beta value (say the Fisher transformed Z scores) would get halved in case of 0.5. If the weights are symmetrical, then the statistic value should not change. For example, say you were only interested in the simple main effects of a single condition in a single group. You would be selecting that particular subject group and condition and simply enter 1 into the Conn interface. Let's call this beta value beta_1. If you repeat it for a different group, you would get another beta value, beta_2. Now, if you are testing group 1 > group 2, you might enter the weight as [1-1], in which case the resulting beta would be beta_1 - beta_2. Instead if you enter the weight as [0.5 -0.5], the beta would be (beta_1*0.5) - (beta_2*0.5). If the contrasts were [1 1] the resulting beta would be beta_1 + beta_2.

One way you can verify this is by loading the resultsROI_Condition00x.mat file from your first level results folder. For a pair of ROIs, say ROI1 and ROI2:

Beta value for simple main effects of the condition for group 1: mean(Z(1,2,1:group_1_length)
Beta value for simple main effects of the condition for group 2: mean(Z(1,2,group_1_length+1:group_2_length)

You can check the values obtained by addition and subtraction and check them with the Conn GUI display of beta. As before, if you have given equal weights to both, the statistic value should remain same.


Hope this helps

Best
Pravesh

Originally posted by Heidi Jacobs:

Dear Heidi,

That is likely related to "centering" of those covariates. The short answer is that you just need to center/demean your covariates. The longer answer is: If your second-level model is defined by selecting 'AllSubjects' and 'education' and entering a [1 0] contrast, this is evaluating the average connectivity in your sample at the zero-level of your education covariate (imagine a regression line between education in the x-axis and connectivity values in the y-axis, you are trying to estimate the height of this regression line when it intersects the y-axis; i.e. the connectivity level at the zero-level of your 'education' covariate). While this may make sense for some covariates where the 0-value has some particular meaning of interest, in most cases one often wants to center the covariates so that the new 0-level of the centered covariates simply represents the average-level/center of your group. In practice, in CONN, simply go to Setup.Covariates.SecondLevel, duplicate your 'eduaction' covariate, rename it to 'education_centered' and then either: a) enter in the 'values' field the string "education-mean(education)" without quotes; or b) select covariate tools 'orthogonalize' option, and click on 'AllSubjects' and 'Ok' (orthogonalizing wrt AllSubjects is the same as centering)


Regarding QA_GCOR that is an automatic Quality Assurance covariate computed during the Denoising step. It represents the average global correlation (GCOR) for each subject and each condition (i.e. the center of the voxel-to-voxel histograms that are displayed in the Denoising tab) which has been proposed/used in several studies as a control covariate when comparing groups of subjects across different sites, or when needing additional control for potential noise/artifactual effects on your connectivity measures. 

Hope this helps
Alfonso

Originally posted by Heidi Jacobs:

Dear Alfonso,

Thank you very much!! This is indeed makes a difference!
Just another thing that I bumped into... 
- In the past I ran a masked ICA by choosing a mask in the setup (and limiting the ICA components to 6)
- now I am rerunning this data with the brainmask of MNI and trying to run a normal ICA, but for some reason it keeps on giving me 6 components.
The ICA window in the single-level analyses is also not complete (see attachment).
Any thoughts on this?

Thanks
Heidi

I solved it! No worries :)