Dear list,
I have problems with modelling BOLD responses in an sustained selective
attention task and I would be very glad if somebody could check my
current model and give me some advice on how to improve it. I apologize
for the lengthy description.

I will first explain the general design of the study: We are interested
in brain regions contributing to successful speaker separation in
challenging listening situations. We set up a task in which participants
had to listen selectively to one of two competing speakers. This
dichotic listening blocks had a duration of 60 seconds. They were
randomly interspersed by blocks of 20 seconds duration, in which either
one speaker only was presented or a question on the content of the
attended story was asked. The TR was 2 seconds. The total duration was
35 minutes and contained 24 dichotic listening blocks. We simultaneously
acquired EEG during the experiment, which provided us with a marker of
stimulus extraction/separation in each dichotic listening block. The
idea was to use this EEG marker as parametric modulator of activity in
the dichotic listening blocks to learn about the relationship between
BOLD responses and successful speaker separation.

I'm aware that the 60 second block length is not really suitable for
analysing BOLD responses, but we had use this block duration in order to
reliably extract our EEG marker from each block. We piloted the
experiment with shorter block lengths and the EEG was just too noisy, so
we now somehow try to cope with this restriction.

The first fMRI model was just a simple GLM containing three task
regressors (single speaker, question, dichotic listening). The high pass
filter was adjusted to 360 sec (i.e. two times the maximum distance
between two successive dichotic listening blocks). However, during
dichotic listening this analysis did not reveal much more than auditory
cortex.

We then thought that using a FIR model may improve the modelling of the
dichotic listening blocks, since not all brain regions may be active all
the time but may rather fluctuate over time. So I set up a model
containing 35 regressors modelling the BOLD time course in the dichotic
listening blocks (i.e., all 30 volumes during the block + 5 subsequent
volumes). The other two conditions were included as standard HRF
regressors. A within-subject ANOVA on the group level, containing the 35
FIR regressors, now shows the rather expected pattern of activity,
including auditory cortex, several frontoparietal regions, the insula,
and subcortical structures (i.e., the main effect of time bin). Beta
time courses extracted from these regions suggest that indeed several
regions are only active in response to trial onset/offset. However, I am
not really sure whether this design is valid at all.

1. Can I mix standard HRF and FIR regressors into one model? Or should I
rather use FIR regressors only? In this case, I would have to use
different numbers of boxcar regressors for each condition, though,
accounting for the different block length.

2. When I searched the web for FIR models I found different versions.
When set up using the SPM GUI the FIR model contains an additional
boxcar regressor spanning the whole duration of the trial. Other people
however seem to neglect this "trial constant" and just use regressors
for each volume. Are both versions valid?

3. Can I include a parametric modulator into my FIR model by just adding
another set of regressors which are multiplied with (the mean-centred)
EEG values? If so, how can I analyse this data on group level? Using an
ANOVA as well?

4. Are there other/easier options I could try out to model my long
stimulation blocks?

Thanks a lot for your help!
Sebastian Puschmann

--
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