Dear Dr Polak,

Here's an ugly solution.

You presumably have prior expectations about where your significant 
voxels are going to be.  In this case, you can just specify your 
events correctly for this brain slice.  E.g. assuming your top slice 
was acquired first, if you expect activations in your bottom slice 
then you need to subtract 0.96 (2.76/2.88) from all of your event 
timings.  Obviously it is now possible to have a negative event 
timing, e.g. an event at the onset of the first scan occurs 2.76 sec 
before the scan reaches the bottom slice, and should therefore have 
an onset of -0.96 scans.

If you don't have prior expectations about where your significant 
voxels are going to be, then you could I suppose do a separate 
analysis for every level in the brain (not every single slice - that 
would be excessive - but maybe predefine 10 levels).  Now you are 
strict with yourself about only looking at the analysis which 
corresponds to the slice that the voxel you are interested in falls 
closest to.  Thus in theory you don't have to make any correction for 
multiple comparisons here, because you are only looking at one 
analysis for each voxel.

Having done this, you could run an analysis with slice timing, and 
see how the statistics for each cluster compares.  You may find more 
power left in the analysis than you expected (OK, I admit it, I 
haven't read Rik's abstract yet).

By the way, you could just choose to include the temporal derivative. 
You can still do a 1 -1 't' contrast on the hrf covariates, to find 
out about the relative heights of the responses to new vs old words, 
and a 1 -1 contrast on the temporal derivative covariates to find out 
about the relative delays between the responses to new and old words.

However, this analysis is making an assumption that the wave form of 
the real haemodynamic response is the same in both cases, with just a 
variable delay.  If this assumption is wrong, and in fact the form of 
the responses to the two conditions is different, then you might be 
introducing a bias.  To take an absurdly extreme example, imagine 
that the BOLD response to new words has a waveform which looks like 
the hrf, whilst the response to old word has a waveform which 
actually looks like the temporal derivative rather than the hrf 
itself.  Even if the response to old words is much larger than the 
response to new words, the voxel could still show up in the 
comparison new vs old.

If you are happy to identify areas which are sensitive to the 
difference between new and old words, but you don't really mind in 
which direction this difference is manifested, you could do an F 
contrast.  I'm not absolutely sure how you implement it; I think that 
it would be an F contrast with 1 0 0 0, 0 1 0 0 (for the hrf and the 
deriv for new words) and 0 0 -1 0, 0 0 0 -1 (for the hrf and deriv of 
the old words).  I think that this gives you voxels where the two 
'new words' covariates explain significantly more of the variance 
than the two 'old words' covariates.  It certainly doesn't tell you 
the direction of any effects (e.g. if the response to new words is 
positive or negative),

Good luck,

Richard.

>Hi,
>
>I am analysing a word reading/recognition task which was run in a 
>blocked design (20 s on, 20 s rest). That is fine and simple.
>
>Within each 20 s on period, there are 10 words presented (1 every 2 
>seconds), half are previously presented words and half are new 
>words. I want to use an event-related analysis to find the 
>difference between new words and previously presented words. This is 
>essentially a stochastic design (since words are presented in random 
>order) with SOAmin of 2 seconds which, according to Friston et al 
>(1999, NeuroImage 10:607-619), should have high efficiency for 
>analysing the differential response.
>
>But I have 24 slices with TR=2.88s (120 ms per slice) so there is a 
>2.76s delay between acquisition of the first and last slice. I can 
>not use the slice timing correction in this case because the 
>frequency of BOLD changes I want to model is much higher than 
>1/{2TR} (according to the abstract of Henson et al, HBM1999, 
>NeuroImage 9(2):S125).
>
>If I include temporal derivatives with the canonical HRF to 
>compensate for this delay between slices, is this an appropriate 
>method? If so, how do I specify contrasts? Does anyone have any 
>other suggestions?
>
>
>_________________________________________________________________________
>Get Your Private, Free E-mail from MSN Hotmail at http://www.hotmail.com.
>
>Share information about yourself, create your own public profile at 
>http://profiles.msn.com.

-- 
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]


%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%