Dear Kalina,

    The mailbase replies are a little sparse this week as most people are at HBM.
So in lieu of a more definitive answer:

> Dear list,
>
> We are trying to specify a subject-specific HRF as a basis function at the
> individual level of analysis. The 'subject-specific HRF' is an
> independently collected and averaged timecourse of signal intensity in a
> cluster in the motor cortex during a separate (motor) task.
>
> It appears that this can easily be added as an option in spm_get_bf.m, by
> adding a 'user specified hrf' option for the Cov variable around Line 91,
> and inserting a condition to import a user-specified vector for bf,
> instead the default SPM's hrf around Line 150, which is now:
> [bf p] = spm_hrf(dt);
>
> However, I was wondering, are there any properties that a Gamma
> basis function has (and that a user-specified vector may not have)
> that are used at some later stage of the statistical modelling and the
> benefit of which would be lost in the case of a user-specified vector?
>

    I can't think of any properties that a gamma bf has that would influence the
statistics of your analysis above and beyond being a better or worse fit to your
actual data. A single gamma bf just happens to be a shape that looks 'hrf-ey' and
has the added advantage of being described completely by a single parameter. The
more mathematically gifted may wish to correct me if I've over-simplified things
here, or just got them plain wrong!

    As far as I am aware, the main problem with using any single parameter basis
function to describe a complex waveform such as the hrf is one common to all
attempts to fit a model to data in linear regression - the model may not be well
specified. You can't fit a square peg into a round hole, and so it is often the
case that the choice of bf is not appropriate. There are options in SPM that take
this into account, and allow the modelling of neurovascular responses by a basis
set of more that one function (either the Fourier or 3 gamma bfs options). These
will fit any example of the 'family' of responses that can be described by a
linear combination of your bfs. The disadvantage is that it becomes harder to
relate these more complex fits back to the underlying neural activity that we
assume generates the hrf, and so unambiguously talk about differences between
evoked responses that we wish to describe by a difference in parameter estimates.

    A recent study by Geoff Aguirre and colleagues showed a great deal of shape
difference between the hrfs of different subjects in the region of the central
sulcus to a transient motor response (The variability of human, BOLD hemodynamic
responses; Neuroimage 1998 Nov;8(4):360-9), but less variability in hrf shape
within subjects when studied over a number of different runs/sessions. This group
now regularly defines subject-specific hrfs which are then used in subsequent
analyses - the approach is described in 'Using event-related fMRI to assess
delay-period activity during performance of spatial and nonspatial working memory
tasks. Brain Res Brain Res Protoc 2000 Feb;5(1):57-66'. This seems to suggest
that your strategy of independently defining subject-specific hrfs is a sound
one, assuming you wish to use these to fit responses in the same region, as hrfs
may show spatial variability even in the same brain. An SPM answer just wouldn't
be an SPM answer without a few caveats, would it?

Best,

Dave McGonigle.







>



%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%