general-discussion > BASC: the size of block length in CBB
Showing 1-3 of 3 posts
Results per page:
Oct 15, 2015  06:10 AM | Kangjoo Lee
BASC: the size of block length in CBB
Hi Pierre,

I was recently asked a question about the length of temporal block (h) that has to be determined to replicate data using CBB.

What is suggested from BASC is h=the square root of the total number of time points (T) in each run, in order to preserve the temporal structure in the resting-state fMRI time-series.
So if T=200, h can be randomly selected between 10 and 30, for example.

Assuming another case, when T=500, for example, h is now becoming larger.

The question was, for the "resting state" data, shouldn't a desired level of the temporal structure be the same for any T? 
So why not using a temporal block of h=square root of 200 for data with T=500?

What is the best answer for this question?

Oct 15, 2015  07:10 AM | Kangjoo Lee
RE: BASC: the size of block length in CBB
one more question, can we apply the same criterion for multi-band resting-state fMRI data?
Oct 15, 2015  02:10 PM | Pierre Bellec
RE: BASC: the size of block length in CBB
Dear Kangjoo,

The experiment that originally motivated me was to plot the standard deviation of correlation coefficients across bootstrap samples, as a function of block length. The rationale was to have as pessimistic a view as possible, and to maximize variability. The results of this experiment are presented in (Bellec et al, 2008) Figure 3.1 (see text beginning of section 4.3 for the results, and end of section 2.2. for the procedure). The experiment is actually about differences in correlation coefficients, but I got similar results on the coefficients themselves (not shown). The dependency between h (block length) and the std of correlation values is smooth with a clear maximum. Parameter h does not seem to have a big impact within a reasonable range. So for BASC I decided to randomize it. The sqrt(T) centered interval is more or less arbitrary but was inspired by the curves from the paper. 

I guess you could easily reproduce the experiment for multiband data, and see if the same kind of behaviour is observed. 

I hope this helps,