 |
 |
Dear Friederike, In general you should aim at similar design matrices for your subjects as far as possible, e.g. with three sequences within your experiment, each of them with a randomized presentation A, B, C, D, E. Otherwise conditions might be affected by unspecific effects like fatigue to different extent (e.g. if A is the 1st, 3rd, and 5th block vs. B being the last, 3rd last, 5th last). Concerning the HPF, you could go with something slightly larger than 2x the maximum of your different onset-to-onset intervals. However, you don't have to account for low frequency noise with the HPF implemented in SPM. You could also (and in your case I would prefer to) go with polynomial regressors = regressors accounting for linear, quadratic, ... changes over time. It has been noted that application of a HPF can actually increase the number of false-positives (Smith et al., 2007, Neuroimage), and it has also been argued that HPF perform poorly when signal is present within the slow frequency bands, as you not only remove noise but also stimulus effects (Kay et al., 2008, Human Brain Mapping). In that case it might be better to go with polynomial regressors, in Kay et al. they suggest to go with polynomials of the order 0 to 4 (constant, linear, ..., quartic), although you might sometimes read that most of the noise is linear and quadratic (no idea when/where this was shown though). Note that there are several other papers focusing on low frequency noise, but there seems to be no consensus / optimal solution. For sure, there is lots of low frequency scanner noise which we want to control for. However, in case your signal has rather high energy within low frequency bands it seems to be better not to remove the signal but rather to try to account for typical noise with polynomial regressors. Best, Helmut