Hi Lukas,

You're reasoning is reasonable. For the high pass filter, you want to make sure that you don't exclude any block-related signal differences, so indeed you need to set a cutoff that corresponds to a time somewhat longer than the longest expected block-related fluctuation. If all you blocks were the same, this would be equal to your full cycle length. However, since you have two different types of block, things are a little more complicated. Imagine that an area of the brain only activates for one of the block tasks and not the other. If the two types of block are presented in alternation (with interleaved "baseline" periods), then the signal of interest would have a period of twice the cycle length. In this case it might be prudent to set a high pass filter cutoff corresponding to 3 or 4 full cycles.

For the low pass filter things are a little easier. The hemodynamic response typically has its quickest change at onset, and this part of the curve corresponds to a frequency of approx. 1/6 Hz. Unfortunately, your TR is only 1/3Hz, which means that the Nyquist frequency is 1/6Hz. So low pass filtering with a cutoff of 1/6 Hz will do no good.

It is my observation that if you have sufficient replications of each condition within subjects, then the parameter coefficients (i.e. beta weights or contrasts) will not be much affected by whatever filtering you use. What will be affacted are the significance estimates of those parameter estimates. But those significance estimates are of no relevance if you are entering your parameter estimates into a second level group-wise analysis. Thus in the end, I have found filtering to have only a marginal effect on my results.

Tom