Hi Emily,

There are many opinions about how to process

resting state data. But if you have physiological

recordings, then it would be good to incorporate

them.

The main difference between examples 5c and 11

is that 5c uses bandpassing, while 11 uses tissue

based regression. Either way is reasonable and

defensible. I personally do not like band passing,

but that is up to you.

Here is some of my more formal whining about

the cost of band passing:

One aspect of band passing that may be commonly ignored is its cost in

terms of degrees of freedom (DoF). Projecting out frequencies can be

viewed in terms of a linear regression with sines and cosines at the

relevant frequencies to project out. So the question becomes, how many

DoF remain?

Consider 3 examples, each having 1000 s of data, but at varying

TRs of 2 s (500 time points), 1 s (1000 time points) and 0.1 s

(10000 time points). Focusing only on the low-pass side, assume

the common pass band of 0.1 Hz, or one cycle every 10 s.

At TR=2s, 40% of the DoF remain, because the Nyquist frequency

is 0.25, and 0.1/0.25 = 0.4. So out of 500 time points, 200 remain

and 300 are lost. At TR=1s, only 20% of the DoF remain, as 0.1/0.5

= 0.2. So out of 1000 time points, 200 remain and 800 are lost. And

at TR=0.1s, only 2% of the DoF remain, as 0.1/5 = 0.02. So out of

10000 time points, 200 remain and 9800 are lost.

This demonstrates that when band passing below 0.1, the effective

TR becomes 5 s, leaving only 200 time points for 1000 s of data,

regardless of the scanning TR. At a very fast TR such as 0.1 s, the

cost is staggering. Even at TR=2 s, using 200 regressors to model

one aspect of noise in a 500 time point model is very expensive.

That band passing can be done in a separate step using a Fourier

Transform does not change the cost, it just makes it harder to see.

</bandpass rant>