Hi SPMers,
We are currently struggling to understand the best way to apply a high pass filter to one of our blocked-design fMRI tasks. The design is as follows:
A B1 A B2 A B3 A B4 A
Where A is a control block lasting 38 seconds, and the Bs are the conditions of interest, each lasting 50 seconds. The Bs are all similar (pictures of faces), but with a variation for each block that is also of interest (each block has a different facial expression).
For some time, we have preprocessed the data in SPM using the default highpass filter 128s cutoff, but it’s recently come to our attention that this cutoff may be inappropriate since our blocks are longer than what is typically recommended. We have tried using a 256s cutoff and find that the resulting contrast estimates for the individual B blocks (e.g. B1 > all As) are quite a bit different from those obtained with a 128s cutoff (correlated ~0.65), while the contrast estimates for "all Bs > all As" are very similar (r~.95). Since this has sizable implications for analyses of behavioral associations with individual block contrast estimates, we want to be sure we are making the best possible choice. As we understand it, it’s a trade off of cutting off too many frequencies of interest vs allowing to much low frequency noise to pass.
We have seen recommendations for a cutoff of twice the length of time between onsets of the same condition, which for us comes to 176s when considering all Bs together (so the time between the onsets of B1 and B2, which is 88s), so we are likely to try that next. In any case though, we’re not entirely sure how this relates to being able to look at a single block (e.g. B1) individually (single a single block is composed of more lower frequencies), and in fact are wondering if our task design has completely undermined our ability to do so - e.g. would we have been much better off having two 25s blocks for each of the Bs so that the resulting signal appears at higher frequencies?
The other thing we have noticed is that pairwise correlations between contrast estimates for the individual B blocks (e.g. B1 and B2, B2 and B3, etc) change dramatically depending on the cutoff used. In fact, with a 128s cutoff B1&B2 are negatively correlated (r ~ -.3), whereas with a 256s cutoff they are positively correlated (r ~ .3)! So again we are left wondering whether and how we'd have the ability to study the relationship between these conditions (facial expressions) with our design.
Any insights into this matter would be greatly appreciated!
Thanks!
Annchen Knodt
|