Global Mean Signal as a Nuisance Regressor?

Submitted by kvoigt on Wed, 02/20/2013 - 05:55

 Hi,

I want to perfrom a seed-based functional connectivity analysis. Before that, I want to regress out noise from the voxel time series. I found a clear justification for treating the 6 head motion parameters, CSF and grey matter signals as nuisance regressors. However, based on e.g. Murphy et al. (2009) I am not quite sure whether to put the global mean signal as an additional regressor or not. Given that I want to compare my results to previous studies (which put the Global mean signal as an regressor) I feel almost "forced" to put it in too, despite of the fact that I am convinced that it might induces artificial anticorrelations.

Any opinions regarding that issue? What would you suggest? 

Submitted by ZHANG_RESTadmin on Wed, 02/20/2013 - 06:42

I will say that, from my personal experience, whether or not regressing out global signals did not always influence the functional connectivity result that much. It depends on which seed region you chose. If you chose default mode area as a seed, it probably induces negative correlation; but if you chose another seed, result may not change too much. A test-retest paper (Zuo et al., 2013) tells that regressing out global mean will reduce reliability of ReHo result (another resting-state metrics). There are some papers discussing about it, they are Chai et al., 2012; Fox et al., 2009; Murphy et al., 2009, and Saad et al., 2012. I suggest not include.

Submitted by kvoigt on Thu, 02/21/2013 - 06:58

 Thank you both for the really helpful (!) answer. I will check the references in more detail again in order to justify my final decision. Thanks again for that!

Besides, I have already run the analysis with / without regressing out the GMS before doing a voxel-wise functional connectivity analysis (Seed: Amygdala). For a brief summary see: https://www.dropbox.com/s/zyfbt5338agqztd/FC_Tmaps.pdf .
However, the results are different and I have problems with interpreting it: 

(1) I only get positive correlations with the amygdala
(2) Using FDR (q=0.05) almost everything is connected to the seed region
(3) Only when using a very high threshold (of 5), then some meaningful cluster arise. 

For interpretation, I hardly can image that I can assume that there are no inverse couplings with the amygdala at rest (given that Roy et al., 2009 does not say this as well (but they used global mean signal regression...))?  Or how can I interpret it in a correct way? Or did I do anything wrong?

I really much value your opinion on this specific example.
 

Best wishes,

KV

 

Submitted by YAN Chao-Gan on Sun, 02/24/2013 - 16:44

Hi,

This is quite normal if you don't regress out the global signal - almost all the brain are highly correlated! You can refer to Fox et al., 2009 to see more details about the anatomical specifity issue and see the controvercial from Murphy et al., 2009. In both papers, you will find the one-sample t-tests without GSR is similar to what you just demonstrated.

Best,

Chao-Gan

Submitted by YAN Chao-Gan on Wed, 02/20/2013 - 10:56

Hi,

Global singal regression (GSR) is a very controversial issue in preprocessing. Murphy et al. (2009) and Saad et al. (2012) oppose to use GSR, while Fox et al. (2009) support its usage, found GSR increases anatomical specificity.

GSR introduces negative correlations, some people do GSR, but then only focused on the positive correlations. GSR also helps to reduce motion effects (Satterthwaite et al., 2013 and a recent ongoing work).

If you don't do GSR, then usually reviewers will be fine. In the current case (if want to compare with those previous studies), you can go with GSR, but put the results without GSR as supplementary. Hopefully you will get similar results for group difference.

Best,

Chao-Gan