Regarding still finding correlations in the ventricles, that is not a surprise. If you computed principle components within a ventricle mask, there would probably be 3-5 major signals, possibly including some BOLD. So regressing out a mean ventricle signal will not prevent the possibility of correlations within the ventricles (unless all of the ventricles had just a single signal).

Regarding bandpassing in your case, note that you will likely wipe out most of the data with that. To be specific, with a short TR=0.72, the Nyquist frequency is 1/1.44 = 0.694444 (the highest frequency detectable in your data). Bandpassing below 0.08 means keeping approximately 0.08/0.69444 = 11.5% of the frequencies. With the slow components and motion, that means your projection matrix will have about 600 regressors of no interest to project from a 667 time point time series. It might be good to ponder that a bit.

Yes, the updated script looks better. Though heavy bandpassing with such a short TR might have its own consequences.

- rick