About Granger Causality Analysis

Submitted by YukiSakai on
Dear REST experts,

 Thank you for your excellent software REST and DPARSF.
 I'm interested in Granger Causality Analysis (GCA). I think that shorter TR is recommenced when we try to do GCA.  When we use REST GCA tool, how long TR will it be recommended?

 Thank you in advance,

Yours sincerely
 

Yuki Sakai


Dear Mrs. / Miss. Sakai:
I think  it's difficult to say what TR should be recommended. It's all depends on your study.
For GCA, the TR is correlated with the order (how many time points to lag to build the model). There are some methods (AIC/BIC) to
help you to set the order according to your TR and your time series.
In my opinion, the advantage of short TR for GCA is that it is more apropriate for neuro dynamic system. But in this case, the
GCA model will be made too comlicated to explain. The advantage of long TR for GCA is that your GCA model will be
much more simple and the results are easy to explain.
I hope it can help.
Yours sincerely
Zhen-Xiang Zang.

Dear Sakai,
Some time-resolved fMRI studies have found that the time delay between brain areas may be about a few hundred milliseconds. So, short TR may be helpful to reveal the Granger causality. However, to my knowledge, there has been no experimental study to show how long TR is better.

Yufeng

 Dear REST experts, 
 
Thank you very much.
 You mention some methods (AIC/BIC) which are used to set the order according to TR and  time series. I am sorry, but I am not sure about those methods. Is it possible to tell me the recommended order when I use 2 sec as TR.
 Thank you in advance, 

Yours sincerely
  

Yuki Sakai

From my personal view, an order of 1 with TR = 2 s implies that the timeseries A is 2s earlier than the timeseries B. During my personal communication with Professor Mingzhou Ding from University of Florida, he told me that I can even try order of 5 (when TR = 2 s) if a low-pass filter is applied. One may not believe there could be 2 s or longer delay between brain regions. But who knows during a complicated resting state? So I will suggest you to try any order in part of your data. Then, using another part of data to  validate the optimized order.
Forums