Dear Valeria,
>Hi all, we have some questions here about how to model a simulated data
>set. We basically have a time series of fixed SOA with only signal
>versus noise scans. The first 4 scans are baseline noise + signal (X +
>(3*SD)) and the following 8 scans are just baseline noise with no
>signal. We tried analysing this dataset with SPM and depending on how we
>model it we obtain very different t values. Analysis I= we model only
>the stimulus (as an event) and ignore the baseline (i.e. we only enter
>one condition (contrast = 1 )); here we found that the t statistic
>ranged between 11 and 5 for the simulated active regions. Analysis II =
>we model both the stimulus (as an event) and the baseline (as a box car
>function); here the t statistic for the same active areas ranged between
>3.21 and 1.6. The degrees of freedom were practically the same for both
>types of analyses (92.4 and 91.7, respectively). Since our baseline is
>not really a baseline (cause we generated no activity in it), we thought
>the appropriate way of analysing the dataset was by using analysis I,
>but we are not sure anymore. Given the nature of our dataset what should
>be the most appropriate way to analyse it, using analysis I or analysis
>II?
You haven't been very specific about what your 'stimulus' is. Is
this a train of events? Or just one event at the start of your 4
'noise+signal' scans? Also, have you convolved both the design
matrix and the simulated data with the hrf (as you would need to do
if you want it to look anything like real fMRI data)? If not, what
do your 'events' actually look like? How long are they?
Anyway, I think that the answer to your question is basically going
to be the same.
A train of events during the first four scans (or even a single event
at the start if the TR is quite short and you have convolved with the
hrf) yields a covariate which looks quite like a box-car for the
first four scans (i.e. the box-car and the event-train would have
quite a bit of shared variance). If you model both the event train
and a baseline shift (box-car) in the design matrix, then the real
shared variance is partitioned between the two in an arbitrary way,
determined by the noise. SPM will probably show you that the
parameters are not uniquely estimable (i.e. you'll have some grey in
the bar at the bottom of the design matrix). In this situation a
contrast of '1, 0', testing a single beta, is not really valid.
Using a contrast of '1, 1' is OK, and would probably give you a
result which is fairly similar to the one you obtained for the
contrast '1' in analysis I. If you only include the event-train in
the model, as in analysis I, then all of this shared variance is now
explained by the event-train, and the beta is correspondingly greater
and more significant than either of the individual betas in analysis
II.
For simulated data, there isn't really a 'more appropriate' way to
analyze the data, although in your analysis II, using the contrast
'1, 0' is always going to be wrong. In a real experiment, you would
usually only model the difference between the stimulus conditions and
the baseline conditions once, to avoid having a rank-deficient design
matrix (i.e. one in which the columns are not independent).
However, one could imagine a situation in which you expect a change
in baseline between the two conditions, which doesn't interest you,
and then an event-related response over and above this which does
interest you. In this case, you would want the shared variance to be
modelled as a confound (because it might be attributable to the
uninteresting baseline shift), so you would need to leave the
baseline shift box-car in the design matrix, as in analysis II, but
have a model of the event train which has been orthogonalized with
respect to this baseline. This latter regressor will only have a
significant beta only in brain areas where there is a significant
response to the events over and above the baseline shift. This is a
very stringent test, though, and in many circumstances no voxels
would exceed a reasonable statistical threshold.
Best of luck,
Richard.
--
from: Dr Richard Perry,
Clinical Lecturer, Wellcome Department of Cognitive Neurology,
Institute of Neurology, Darwin Building, University College London,
Gower Street, London WC1E 6BT.
Tel: 0207 679 2187; e mail: [log in to unmask]
|