Dear Erik,
Although I am by no means an SPM expert, I thought that I would communicate
some thoughts. I wrote most of this message before I saw that Karl Friston
has already written a reply, and since mine is not entirely consistent with
his, you may well choose to ignore mine!
As you have identified, there is a theoretical difficulty in interpreting a
study with a design matrix which includes a row which models a number of
individual events within a condition, and another row which models a change
in the 'baseline' signal level during that condition. The problem is that
the two are not independent.
The extreme example is, of course, any standard epoch-based design
comparing events. Thus, if we present lots of stimuli of type A, all very
close together, in one block, and lots of stimuli of type B in another
block, the design matrix might contain a row, to model the effect of
interest, which steps up from zero to one at the start of this block, and
back down from one to zero at the end.
When there are many events close together, then, a row treating them as a
block and a row treating them as individual events may be almost
equivalent, and if both rows exist in the design matrix, then the extent to
which each is used to the model the data may be determined largely by noise.
If the analysis (with 'block' and 'events' rows) yields a parameter
estimate for the 'events' row which is significantly different from zero,
then can one safely say that there is a change in the 'event-related
response amplitude' during condition A with respect to condition B? Not
necessarily: the program may have been using this row in its attempt to
model what is in fact a change in 'baseline' activity. You would need to
convince the scientific community that your events are sufficiently sparse,
that the 'block' row and the 'events' rows are not confounded.
For this reason, I don't think that Solution 1 is a possibility, because
even if the SPM{F} shows that the data is significantly influenced by this
one event type, you cannot be sure that this is not because it so happens
that the 'block' row doesn't fit the change in baseline very much better
than the 'event 1' row.
Solution 2 is also not possible: certainly if a particular voxel came out
as significant in all three of your contrasts (1, 0, 0; 0, 1, 0 and 0, 0,
1), then you might begin to suspect that this was because a 'baseline
shift' was being modelled partly by your 'event' rows (even if it is
'mostly' modelled by the 'block' row). In fact some voxels may show up as
significant for some events but not others just because of noise.
Solution 3, though, seems to give interpretable results. Now you are
empirically testing the hypothesis that a given 'events' row provides a
better fit of the data than the 'block' row. However, in this case it is
clearly going to be impossible to distinguish voxels in which there is a
significantly greater positive 'event-related' response in condition A than
in condition B from those in which there is a significant negative baseline
shift during condition A. This seems to me to be a rather uncomfortable
ambiguity.
For this reason, I would promote another solution, solution 4. This is to
specify all three events and the block as effects of interest (as in
solution 3), but in addition to comparing each event with the block (eg. 1
0 0 -1), to also compare events with eachother (eg. 1 -1 0 0). Ideally
there should be an equal number of each event type during the experiment,
so that provided one can assume that the time-course of the response to
each event is approximately the same, no one event will be more confounded
with the 'block' row than any other event. If during condition A the
ordering of event types is randomised, then these comparisons seem to me to
be safe.
Best wishes,
Richard Perry.
> Dear spm,
>
>I have a question concerning the specification of contrasts in an
>event-related analysis.
>
>The baseline of my experiment consists of a blocked design, alternating
>between the conditions A(active) and B(rest).
>During condition A, several short changes in the (visual) stimulus occur.
>I have implemted these in SPM97 as single events (hrf-function) and an
>impulsetrain (condition A).
>I'm interested in the single events per se, i.e. those regions that are
>extra activated or deactivated compared to the
>activationlevel during condition A.
>
>I see three possible solutions:
>
> 1. I look at the SPMF for one event-type, having implemented all other
>events and the block as CNI
> 2. I define the block as CNI and all event types as COI; I define the
>contrasts(for three event types) [1 0 0; 0 1 0; 0 0 1]
> to look at each event type individually.
> 3. all event types and the blocked covariate are implemented as COI
>(e1 e2 e3 block).
> Then I define the contrasts [ 1 0 0 -1;
> -1 0 0 1;
> 0 1 0 -1;
> 0 -1 0 1;
> 0 0 1 -1;
> 0 0 -1 1]
> to look at the activations or deactivations compared to the
>blocked respons.
>
>
>My questions:
>
>1) What could be the benefit of using a contrast [1 0 0] instead of the SPMF?
>2) As I see it, the implementation of the blocked covariate as COI or CNI
>wouldn't influence the beta's, but would
>change the z-maps, because of its influence on the effective degrees of
>freedom. So, if I'm only interested in the the
>differential respons in between event types, would it be better to
>implement the blocked covariate as CNI? In general,
>what's the influence of an extra COI on the SPMZ's?
>3) If one is interested in the extra activation caused by a single event
>during an active condition A, should one look at the
>F-map (with the block as CNI) or is the contrast between the event type
>and the block a more appropiate, valid way?
>4) Is there another solution to this problem?
>
>
>Thanks in advance,
>
> Erik
>
>
>
>~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> Erik Beatse
> Department of Radiology, KUL MR Research Centre
> University Hospital Gasthuisberg, K.U.Leuven
> Herestraat 49, B-3000 Leuven (Belgium)
> Phone: +32 16 347753, Fax : +32 16 343765
> <mailto:[log in to unmask]>mailto:[log in to unmask]
>~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
from: Dr Richard Perry BM BCh MA PhD MRCP(UK),
Clinical Research Fellow, Wellcome Department of Cognitive Neurology,
Darwin Building, University College London, Gower Street, London WC1E 6BT.
Tel: 0171 504 2187; e mail: [log in to unmask]
Pager: 04325 253 566.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|