I'm just starting to use SPM and I'm trying to compare its methods to things
I've been doing with other software. I'm fairly new to this list, but I've
searched the archives and can't to find answers to these questions.
My questions relate to estimating and orthogonalizing the temporal
derivative of the HRF.
1. SPM2 (in spm_get_bf.m) estimates the temporal derivative by subtracting
time points that are one second apart regardless of TR. What is the logic of
hardcoding this value? This should be perfectly fine for TRs that are
approximately one second or larger, but couldn't this be detrimental for
significantly shorter TRs?
2. Going through the SPM mailing list archives, there is some discussion
about how the a well-behaved function and it's first derivative are
orthogonal, but filtering ruins this orthogonality. Therefore, SPM
orthogonalizes the basis functions. In spm_get_bf.m, SPM orthogonalizes a
high temporal resolution version of the HRF and it's derivative (sampling
at 0.0625s). The regressor in the design matrix uses basis functions
downsampled to the TR (calculated in spm_fMRI_design.m). Since downsampling
is essentially a filter, why aren't the downsampled basis functions
orthogonalized?
3. On a similar note, what is the purpose of orthogonalizing the basis
functions? From what I understand, orthogonalization is useful because it
keeps regressors independant, but unless one uses a very widely-spaced
design, orthogonal basis functions convolved with a reference function are
not necessarily orthogonal. I looked at regressors in SPM.xX.X from my
widely-spaced design (stimuli every 18-22s) and they are definitely not
orthogonal. Is this design matrix orthogonalized at some later step before
parameter estimation?
Thank you for you help.
Daniel Handwerker
Helen Wills Neuroscience Institute
University of California, Berkeley
[log in to unmask]
|