I hope someone could give me some advice regarding high-pass filtering in sparse-sampling designs. I have checked several websites on this but couldn't find a satisfying answer.
Our sparse sampling paradigm involves three conditions. "On" conditions a) and b) containing 36 acquisitions each, plus 18 silent baseline or "off" conditions. We concatenate always six conditions of the same type. These blocks are alternating throughout the experiment. Our TR of 10s includes a constant 7.8s delay, in which vocal tasks are performed.
When I do my first-level analyses, SPM8's default HPF is set at 128s. As far as I understand, the rationale for this default value is based on standard block designs. I read though that data acquired with sparse-imaging designs "will induce changes in BOLD activation at a lower frequency than an equivalent continuous imaging design. Also, because the time-series is sampled much less often in sparse imaging, the nyquist limit will be correspondingly lower and therefore there will be much less high-frequency information in the time-series (Matt Davis, posted Thu, 3 Feb 2005). Thus the filter settings must be different.
What confused me was this observation. When I use the 128s filter, with a FIR basic function (10, Order 1), half of my signal of interest in the motor task is removed. Exploring my design with SPM8 revealed that this HPF cut off half of my largest frequency density slope (between 0.005 - 0.01). There's not much going on above that frequency. A smaller peak is found between 0.001 and 0.005 (approx.). Higher filter settings (e.g. 239s or 350s) seem to leave my activation of interest intact.
So I started looking more closely into this issue. According to the Nyquist Sampling Theorem, removing signal fluctuations that are greater than half the frequency of the task potentially removes signal of interest. Thus, the cut-off frequency should be half the task frequency. However, this might not apply to sparse-sampling as Matt Davis suggests in his post from 2005. He argues that not applying an HPF for sparse sampling designs would be an option too (https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind05&L=SPM&P=R64238&1=SPM&9=A&I=-3&J=on&X=745F7E61DF6160659E&Y=boris.kleber%40uni-tuebingen.de&d=No+Match%3BMatch%3BMatches&z=4). I was wondering what other people thought about that?
The combined time of the on- and off-blocks times 2 is often used as cut-off for a block design. When I apply this approach to my design, one entire cycle of all three condition blocks would be approx. 3min (taking the delay in TR into account). So I'd end up with a 360s HPF, if I am not mistaken. Is this a viable way of defining the HPF for my first-level analysis or should I rather use the graph provided when exploring the design and adjust accordingly?
I know that setting the HPF is somewhat arbitrary but it would be nice to come up with more elegant way of defining the filter. Especially since it has such a great impact on my data.
Many thanks in advance,
Boris A. Kleber (PhD, M.S.)
Institute of Medical Psychology and
Director: Prof. N. Birbaumer
University of Tübingen
Fax.: +49-7071 -29-5956
[log in to unmask]