I write to you regarding the
issue of clustered-temporal acquisition analysis in spm5 and I would very
much appreciate if you could take a few minutes to read this
In auditory research clustered-temporal acquisition this is a big issue
and but we are still unsure about the correct way of analysis.
(Clustered-temporal acquisition would e.g. mean to present the stimulus
every 10 seconds in silence for 3.5 seconds and acquire three volumes (TR
1) approximately 5 seconds pSO)
In earlier studies the box-car function convolved with the hrf was used
to analyse each clustered temporal acquisition as an epoch.
In spm5 we assumed that the FIR-function would be the most appropriate to
use. We then entered 10 seconds as interscan interval, the FIR function
as basis function with window length 3 (three volumes, TR 1) and first
To correct for different T1 between acquisitions we included to
covariates in the model.
However- when using the hrf with time derivatives the results look
approximately the same, even better, even though I can no longer specify
the duration of the acquisition and spm would assume that one volume
acquisition lasts 10 seconds.
Here are my questions:
1) assuming that spm works with a TR of 10 seconds and models the hrf as
in the above mentioned case. Would you believe that this way of modelling
is appropriate, or is that almost like improving the data artificially? I
assume the shorter the inter-stimulus-interval the more suitable the hrf
would be. Is that correct?
Yes; that is correct. The form of the HRF becomes a more useful
constraint when the TR is
small in relation to the temporal extent of the HRF. With a TR of
10 seconds the decimated
regressors are not really affected by the shape of the HRF and there is
little point in using it.
If I understand your design properly, you have a difficult problem
because the TR is not constant:
I.e., you acquire three volumes with a short TR every 10 seconds.
This means the three volumes
sample the response over short periods of time. The most comprehensive
model would be
one which uses a parameter for each of these observations (i.e., your FIR
model). Note that
you should get the same result for any TR here because you are simply
modelling the response in terms
of the a signal on three occasions after a stimulus. It may be that a
better model for these three
responses is their average and their slope - this is effectively the
model you are getting with an
HRF and its time derivative. The final model would be just the average
(you would specify this as
a covariate of interest - without convolution, with three '1's for the
three scans after each event type).
In short, the convolution in SPM is really irrelevant for you because SPM
thinks your TR is ten seconds.
This means you might want to specify your regressors directly as
covariates of interest that can be
one, two or three separate columns for each event type, depending on
whether you want a simple or
more complicated model (with a mean, mean and slope or mean, slope and
2) Is the FIR function as the
"most general" function in spm5 comparable to a box-car
function or would it at least be able to catch modulations of a box-car
FIR and box-cars are models of the underlying stimulus function or
neuronal response. The HRF maps
the neuronal response to a hemodynamic response. In your case the
sparse and non-stationary; so the concept of a convolution kernel (i.e.,
HRF is not really relevent
3) Is there possibly an
alternative way of doing the individual model, which we haven't thought
Try using covariates of interest as described above.
4) does the specified microtime
resolution affect my clustered temporal acquisition
Mathematically, it does; but because you have sparse sampling the shape
of the HRF is largely
irrelevant and therefore it does not really matter what the micro time
resolution of the underlying
stimulus function is.
I hop this helps - Karl