> The activity measured is a weighted sum of the present activity and the
> activity of that voxel from the recent past (done by the HRF convolving of
> the contrast). This HRF is based on the haemodynamic delay (HD) (let's say
> 8 sec) and I think also on the TR (let's say 3.2 sec). The kernel used to
> form the HRF must thus be 8/3.2, because the HD is in seconds, while our
> data is in image-units (1 such unit = TR) and the convolution takes place
> on "something" in image-units. Is this reasoning correct ?
Not quite. The hrf convolution is first calculated in "high-resolution"
time, with default units of TR/16 (set by variable fMRI_T). This
is then down-sampled every TR to generate the covariates (the particular
high resolution time-bin assigned to each scan depending on the variable
fMRI_T0, with a default of 1, ie start of scan).
> One single scan thus lasts for TR seconds, resulting in a difference in
> time between the first and the last voxel of that scan recorded. Does SPM
> correct for this fact ?
You are correct that there can be an important timing difference between
the first and last slice acquired - this is known as the slice-timing
One solution is to interpolate the data in time (the "slice-timing"
in SPM preprocessing). Search the SPM email archive for more details on
Another solution is to use a more general basis set that can accommodate
timing differences. These solutions are discussed in the HBM abstract:
Henson et al
(1999), Neuroimage, 9, 125.
> If yes, does it then presume an ordering in the scan direction (inferior
> to superior, posterior to anterior, ...) and adapts it thus the HRF to the
> spatial location of the voxel. We use sometimes interleaved scanning
> techniques (all slices are not recorded sequentially from inferior to
> superior), resulting in hashed images, which need to be re-hashed to get
> correct images in SPM, meanwhile losing all the temporal information,
> which will certainly result in bad interpretations, conclusions.
Search the SPM email archive for "slice-timing" for more details on
> If not, will the used -uncorrected thus- strategy not lead to wrong
> conclusions ? I can imagine that some voxels, having a nearly significant
> significance level, are abusively marked as significant or insignificant,
> resulting in smaller or larger brain regions than it should be in reality,
> due to this neglected difference in time.
> Or doesn't SPM count for it, but is it anyway negligible ?
By ignoring this problem in models where timing is important (eg
event-related analyses), it can have serious effects, yes (less so
for long blocked designs). See the Henson et al ref above for an
The longer the TR, the bigger the problem.
DR R HENSON EMAIL [log in to unmask]
Wellcome Department of
Cognitive Neurology TEL (work1) +44 (0)20 7833 7483
12 Queen Square TEL (work2) +44 (0)20 7833 7472
London, WC1N 3BG FAX +44 (0)20 7813 1420