Dear Kalina,
> i'm writing with yet another high-pass filter question. What is the
> rationale for choosing a cut-off period for event-related studies where
> each individual event is modelled by a separate HRF function? We have
> such an ER-study in which the trials can belong to one of 3 conditions
> (say a, b, and c). The 3 trial-types are presented pseudo-randomly, and
> there are 24 trials per session, eg:
>
> b a c b c a b c a c b b c a a b c b a c c b a a
>
> the TR is 2 sec and each trial is of 30 sec duration (of which at least
> 15 sec are rest). Using SPM97 and hrf as a basis function, the default
> estimate for a cut-off period seems to be 240 sec, but i'm not entirely
> clear as to what the reasoning behind this estimate is. Thanks for any
> comments here.
The default is simply 2 x min(max(intertrial interval)) over trial
types. This ensures that experimental variance is not modeled by the
confounds. In event-related fMRI some low frequency structure in
modeled responses can be removed as confounds with inpunity and a 64
sec cut-period would be quite justifiable. This is especially the case when you are interested in differential repsonses (which correspond to trial
by time interactions) which cannot be removed by low frequency confounds
(main effects of time). One way to think about this is to imagine the
regressors that ensue after applying the contrast wieghts. Clearly
the difference between a and b has a higher frequency structure than
simply testing for the main effect of b.
Incidently the stochastic nature of your design allows you to use an
inter trial interval that is much shorter than 30 seconds and thereby
increase the efficiency of your design markedly (unless there are other
contraints).
> another question pertaining to high-pass filters in ER studies arose
> when we compared the activation results after two different cut-off
> periods were applied to the same set of data. To my understanding, if
> all other things are kept equal, a shorter cut-off period should result
> in smaller or equal number of pixels significantly above treshold,
> compared to a larger cut-off period. However, in this case, a 60 sec
> filter analysis yielded some additional areas compared to the 240 sec
> filter analysis. How is this possible and is it particular to the ER
> type of analysis?
This is completely possible and motivates the use of a 64 second
cutoff. Your results suggest that there are low frequencies in your
data that are unrelated to the experimental design. Removing them (and
low frequency signal components) leads to smaller errors and greater
sensitivity. It should be noted that too much smoothing and filtering
can compromise the efficiency of estimating the parameters of evoked
responses (particularly low frequency ones, which cannot be estimated
if they are removed). One of the advantages of using temporal basis
functions like the h.r.f. is that all frequencies are estimated jointly
rendering you less sensitive to this loss of efficiency.
With best wishes - Karl
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|