Dear Klaus,
A lot of issues are raised by your question, and I don't think that I
will be able to cover all of them in an entirely comprehensible way.
But I'll do my best.
>this is a question regarding user specified regressors:
>In a standard block design (2 sessions, 5 OFF, 5 ON epochs, 10 scans per
>epoch, TR=3sec) we are interested in differential effects of two types of
>events (SOA=3sec) within the activation and also within the baseline
>conditions.
>
>I thought about using the option user specified regressors to model the
>different event types within each condition as follows:
>
>differential events only in the activation condition:
>0 0 0 0 0 0 0 0 0 0 1 1 2 1 2 2 2 1 2 1 0 0 0 0 ... (a total of 100)
>or
>differential events in activation and baseline condition:
>1 1 2 1 2 2 2 1 2 1 2 2 2 1 1 1 2 1 2 1 1 2 1 1 ... (a total of 100)
>
>In the design matrix I received a third column for the regressor.
>
>If I understand the idea of the regressor correctly, setting the column of
>the regressor to 1 and the other columns to 0 in the contrast manager,
>should detect those voxels that are not activated in the baseline
>condition, show an intermediate increase during the activation period at
>type 1 events and more activation at type 2 events (example 1).
Not exactly. This contrast would detect any voxels in which the
regressor can explain a significant amount of the variance. Such a
voxel may in fact be showing a shift in baseline between the
'activation' epoch and the 'baseline' epoch, without any response to
the events at all. Or it may simply be responding to event type 1.
Or it may be responding to all four events, but the average response
during 'activation' is greater than the average response during
'baseline'.
You are not in any sense testing the hypothesis that certain voxels
are 'not activated in the baseline condition' (indeed I am not quite
clear what this statement means, and it certainly doesn't seem to be
the kind of hypothesis that can ever be supported using 'classical'
statistics). The rest of your statement constitutes at least two
separate hypotheses, that type 1 and type 2 events both produce
positive BOLD responses, and that the response to type 2 events is
greater than the response to type 1 events. These two hypotheses
would need to be tested with at least two appropriate regressors (see
later). Voxels which satisfy both of these might be tested for with
a conjunction.
> In example 2 it should detect those voxels that are activated more
>during type 2
>events compared to type 1 events, irrespective of ON or OFF period.
>
>Is this correct?
Not exactly. SPM99 would mean correct this regressor, so that it
would become something starting -0.5 -0.5 +0.5 -0.5 +0.5 ..... etc.
This would test for voxels which give a significantly greater
response to type 2 events than type 1 events during the course of the
experiment. However, you are NOT explicitly testing the hypothesis
that there is no difference between 'activation' and 'baseline'
conditions in this regard (which is what your 'irrespective' might be
taken as implying).
But let's concentrate on your model.
This doesn't seem to be a very good model for the expected BOLD
response. This is not really what the 'user-specifed regressors'
option is intended to be used for. Normally one would try to set up
this part of the model in the standard 'epoch/events' part of SPM99,
as you suggest later.
>If this is correct, the delay of the hrf would not be taken into account.
>Could this be compensated by adding a 0 at the beginning of the sting vector?
Very crudely, yes. But convolving with the hrf must be a better
option. You don't really expect the BOLD response to look exactly
like the neural response, but just shifted by 3 seconds. Convolving
with a single 'standard' hrf for the whole brain isn't perfect, but
is certainly likely to do a better job.
Another weakness of your present model is that there is an assumption
that your type 2 events produce exactly twice as much BOLD response
as your type 1 events. Any deviation from this assumption will end
up in your error column, reducing the significance of your results
(and adding structure to your error term which strictly speaking
invalidates your inference).
> A different option would be to model the different event-types with an
>event-related hrf-model. This however might be limited by the fact that the
>SOA equals the TR and that there is no other jittering or null-events.
This limitation is not a property of using this particular means of
analysis. By not having any jitter, you are limiting your sampling
resolution of data after each event to your TR, regardless of how you
analyze the data. But given the time-course of the BOLD response,
data sampled at 3 second intervals can still give very useful results.
Modelling the data using 'events' and 'epochs' generated by SPM99
must be a better option. How exactly you do it will depend on your
underlying beliefs about the physiology.
A reasonably comprehensive model would contain all of the following 5
regressor: a 'box-car' regressor for your 'activation' epochs, and 4
event trains for [type 1 events, activation], [type 1 events,
baseline], [type 2 events, activation] and [type 2 events, baseline].
In this model, you are not making any assumptions about the ratio of
BOLD response to events type 1 and 2. You are also allowing for the
possibility of an interaction between type 1 vs type 2 events and
activation vs baseline. For example, it may be that there is no
difference between the responses to events 1 and 2 during the
baseline condition, whereas there is a difference during the
activation condition.
A weakness of this model might be that your events appear to be quite
close together, and as a result, your 'box-car' regressor may be
almost identical to the sum of the two 'activation epoch' event
trains. Certainly these regressors will be far from independent. In
this case, you might want to omit the boxcar regressor entirely and
just model the events.
You may, alternatively, have reasons for believing that the responses
to events type 1 and 2 will not change during the experiment,
although they may be superimposed onto a different background level
of BOLD. In this case the four event regressors could be collapsed
into two. However, you have so many degrees of freedom in most fMRI
experiments that this is very unlikely to be necessary.
>Are there any other options, for example combining the box-car design and
>the specified regressor in certain contrasts?
You can compare them in contrasts if you wish to, but of course in
comparing events with epochs, you are not really comparing like with
like, so the results are not readily interpretable. You probably
have the additional problem that you events and your epochs are very
far from orthogonal. So although comparing the response to type 1
events to type 2 events is perfectly legitimate, you can't really
just look for the main effect of type 1 events, or the main effect of
the 'activation' epoch. If you wanted to do this, you would have to
fiddle about with orthogonalizing regressors. The partitioning of
'shared' variance which can either be modelled by the box-car or by a
linear combination of the the event regressors is arbitrary and will
be determined by noise, which is another reason why comparing events
to epochs is not really useful.
User-specified regressor were really designed for putting in more
complex regressors such as scales for modelling reaction times or
subject's accuracy etc. For simple models in which you have
straightforward events and/or epochs, it is just much easier to build
your design matrix using the 'event/epoch' options. Even if you
didn't want to convolve with the hrf (although it is difficult to see
why you wouldn't!), the option to omit the convolution is offered to
you.
Best wishes,
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]
|