Dear Yifei,

Some thoughts on your questions below

Best
Alfonso
Originally posted by Yifei Zhang:
Yes, you are exactly right regarding the interpretation of the different FDR corrections: "seed-level" correction is correcting across all target-ROIs for each seed-ROI (16 target ROIs), while "analysis-level" correction is correcting across all seed-target pairs (136 pairs). If you do not have any a priori hypothesis about which seed ROIs you would like to test (you would like to test all ROI-to-ROI pairs) then you should use "analysis-level" correction.

Alternatively you can also: a) use a combination of two thresholds, e.g. one "seed ROIs" threshold (e.g. F-test, an omnibus test for each seed ROI) that uses FDR correction across the 17 seeds, and then a "seed-level" threshold (which can use "seed-level" FDR correction, or even uncorrected threshold if only used as a post-hoc test); or b) enable permutation tests, which will allow you to use other forms of "seed-level" thresholds like the one above in order to identify significant seeds, or "network-level" thresholds in order to identify significant networks of ROI-to-ROI connections.

But please see also the answers below regarding the proper second-level contrasts because that applies also to the example that you mentioned here

The second analysis is not correctly defined, it is still including all subjects (MCI and AD as well as HC) and simply leaving their average effects un-modeled. When looking at a single group and in order to additional control for age and gender you could do either one of the following two analyses:

Analysis1:
a) Create two new second-level covariates: "ageHC" and "genderHC" which contain the age and gender values, respectively, for each subject within the HC group, and 0's for the rest of subjects (MCI or AD)

b) center these covariates (or otherwise subtract the desired level of age and gender that you would like to evaluate connectivity at for the HC group; e.g. if you want to evaluate connectivity at age 30 simply subtract 30 from all the HC age values in this covariate, still keeping the MCI and AD subjects as 0's)

c) then in the second-level analysis tab select the "HC", "ageHC", and "genderHC" covariates and enter a contrast [1 0 0]. That will evaluate the average connectivity within the HC group at the chosen age and gender levels (or at the average-age, average-gender levels if simply centering those covariates)

To double-check, your second-level analysis results should report a number of degrees of freedom equal to the number of subjects in the HC group minus the number of "between-subject" effects included (3 in this case)

Analysis2:
A different and perhaps simpler variation of the same analyses would be the following:

a) center the "age" and "gender" covariates (or otherwise subtract the desired level of age and gender to evaluate connectivity at)

b) in the second-level analysis tab select the "HC","MCI","AD","age", and "gender" effects, and enter a contrast [1 0 0 0 0]

To double-check, these second-level analysis results should report a number of degrees of freedom equal to the total number of subjects (in all three groups) minus the number of "between-subject" effects included (5 in this case)

The difference between these two versions of the analyses is that in the former the variance is computed within the HC group only, while in the latter the variance is pooled across all three groups. The results should be relatively similar (effect sizes are identical but stats might differ), the former would typically be referred to as "one-sample t-test controlling for age and gender", while the latter would typically be referred to as "simple main effect of HC" in an ANCOVA model.

Yes, unfortunately both the analysis (a) and (b) that you describe are still not correctly defined. If you use the "analysis1" version mentioned above you should get exactly the same results independently of which groups were included into your CONN project (the "analysis2" version takes into account all of the groups included into your project to build that pooled variance estimate so those results would differ when you have only two vs. when you have three subject groups)
 
Yes, either way is perfectly fine. Masking your ROIs with the subject-specific grey matter voxels is a great way to potentially boost the signal-to-noise ratio in your ROIs but some times it may be problematic in the presence of inaccuracies in the normalization step and/or for relatively small ROIs, so we do not really have a "one-solution-fits-all" recommendation in this regard (it will depend on the particulars of your dataset and choice of ROIs). 

Hope this helps!
Alfonso