There is an old data set lying around the lab that I would like to toy with. Unfortunately, the person who designed the study did not split the acquisition into runs. Thus, there is one 25 minute long acquisition with two alternating conditions in a block design. I am wondering if it is worth the effort to proceed...

It seems like a very bad idea to model a high-order polynomial for the drift terms (e.g., adding a degree every 150 s as has been recommended), but surely the scanner drift is going to have to be modeled somehow. So I have a couple of focused questions:

1. In general, in addition to working against subject fatigue, is it beneficial to split a long acquisition into separate runs (I'm guessing yes), and is the reason to allow the scanner to return to a baseline state between runs? What is the risk of not doing this? Would later acquisitions at the end of the mega-run have poorer SNR than earlier acquisitions? I can't find anything in the literature to suggest splitting long acquisitions into runs, but this is obviously common practice so there must be a reason (or several). A cite would be great if you know of one.

2. Is it appropriate to model this particular 25 min acquisition as separate runs in a regression, even though this was not actually done? My gut says no, but would like to hear other opinions.

3. What is the upper limit on recommended lengths for runs? Maybe this depends in part on the degree of polynomial one is comfortable modeling?

Anthony