Dear Helmut,
Thanks a lot for your detailed response. That helped me a lot!
Indeed, I missed the point that the stimulus duration has to be set to
zero in FIR models, independent of its actual duration. I now
re-calculated the model and included FIR regressors for all three
experimental conditions + the parametric modulation. The 20 second
blocks (single speaker, question) were modelled using 15 FIR bins; for
the 60 sec dichotic listening blocks I used 35 FIR bins.
The results look plausible and are also consistent with my previous
model. However, the statistical power of my effects on the group level
has decreased. I guess this is related to the huge number of regressors
entered in my single-subject FIR model (107 in total!).
The parametric modulation does not yield any significant results,
though. However, this hasn't necessarily to be a failure of the model :-)
Best wishes,
Sebastian
Am 07.01.2016 um 13:59 schrieb H. Nebl:
> Dear Sebastian,
>
>> When I searched the web for FIR models I found different versions
> Not sure what you've found, but it might well (have to) be implemented differently with different softwares. In any case, for a "classical" SPM FIR model the duration would have to be set to 0 s. In fact, there's not meant to be a real duration anyway, as we don't have any particular assumptions about the shape and length (which made us turn to FIR in the first place ;-) ). Rather, a series of stick functions is defined via window length and order (e.g. 30 time bins covering one TR each) to obtain activation estimates for different time bins. Although the stimulus input is 30 s we might want to look at activation patterns for 6, 60 or 600 s.
>
>> When set up using the SPM GUI the FIR model contains an additional boxcar regressor spanning the whole duration of the trial
> If you go with a non-zero duration it's actually not an additional long regressor, but a set of long regressors, each shifted by one time bin, which then undergo serial orthogonalization (independent of the ortho. settings in the GUI, which refers to parametric modulators). The design matrix can indeed *look* like what you observed then, but due to the ortho. and the time shift issues you have to be very careful with regard to proper interpretation of the model. Accordingly, only set up a model like that in case you have very good reasons to do so. If you come across an SPM FIR model with a non-zero duration then most likely, this resulted from people being unaware of the consequences.
>
>> Can I mix standard HRF and FIR regressors into one model?
> This is actually very common when it comes to transient vs. sustained responses. Whether these predictors are reasonable or not is another issue though, as it highly depends on the particular design and your assumptions. See my previous comments at https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=spm;9e20538b.1509 , continued at https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=spm;719e7d0a.1510 . While a certain set of regressors might be justified in one instance (like the combined FIR/canonical HRF in the Visscher study), it might lead to severe confounds in another. E.g. in your case, given the duration of 20 s for the transient event it might be safer to go with FIR predictors for both components, as a canonical regressor (which assumes a constant neural "on" period for 20 s) might already be a rather crude approximation (in contrast to e.g. a brief interval of 2 s). But leaving the timing issues aside, the sustained component doesn't really sound sustained to me, but rather like different auditory input / listening tasks.
>
>> Can I include a parametric modulator into my FIR model by just adding another set of regressors
> Yes. Estimates for parametric modulators can be analyzed just like those for "standard" regressors.
>
>> The first fMRI model was just a simple GLM containing three task regressors (single speaker, question, dichotic listening) [...] However, during dichotic listening this analysis did not reveal much more than auditory cortex. [...] A within-subject ANOVA on the group level, containing the 35 FIR regressors, now shows the rather expected pattern of activity [...] several regions are only active in response to trial onset/offset
> Make sure about your FIR model. If it's based on a non-zero duration, then the interpretation trial onset/offset is likely incorrect, see above.
>
>> Are there other/easier options I could try out to model my long stimulation blocks?
> First of all, are there any baseline / rest conditions / fixation periods with no (auditory) stimulus input? If not, then I would just model "single speaker" and "question", with "dichotic listening" serving as unmodeled, implicit baseline. However, this works properly only if the first "dichotic listening" period is really identical to the second "dichotic listening" period within those 60 s blocks. This might not be the case though, e.g. if "question" refers to material presented before then subjects might attend during the first period but not during the second. Maybe you can provide additional information on the task/design, this should facilitate model choice.
>
> Best
>
> Helmut
>
--
Dr. Sebastian Puschmann
Biological Psychology
Department of Psychology
European Medical School
Carl von Ossietzky Universität
26111 Oldenburg (Germany)
phone: +49-441-798-3931
office: A7-032 (Haarentor campus)
web: www.uni-oldenburg.de/cogneuro
|