Alex -

> We are doing a rapid event related study.  There are initially four
> types of trials.  These are then modified on a subject by subject
> basis by subsequent memory effects.  This leaves us with 8 trial
> types with vectors of onsets that differ for each subject and each
> session.  We initially used the stochastic design feature to make a
> model of the four initial trial types.  We also included fixation
> trials which we did NOT explicitly model thereby giving us some
> "jitter".  The ITI is 3 seconds.  The TR is 1.5 seconds.  Our results
> have been disappointing. I have two questions:
>
> 1) should we explicitly model the fixation trials? i.e. give them a
> column in the design matrix.

You can, but I don't think it will help. For simplicity, assume
you modelled just two event-types (remembered and forgotten):

If by "disappointing" you are refering to the main effect versus
baseline (a [1 1] contrast in this example), then modelling
fixation trials and performing a [1 1 -2] contrast is unlikely to give
better results (possibly worse, given loss of df's). The only situation
where it might improve things is if something is actually happening
in the brain when fixation trials occur (which may be the case if
your fixation trials differ from the interstimulus baseline - ie are
detectable by the subject - in which case, this is not the intended
use of our definition of "null events", which really should be
undetectable
- ie exist simply to give a stochastic distribution of SOAs). In this
case, modelling fixation trials may capture additional structure in the
residuals, reducing error, and improve your results.

If by "disappointing" you are refering to the difference between
remembered and forgotten (a [1 -1] contrast in this example),
modelling fixation trials is also unlikely to affect results much,
assuming a near random distribution of these event-types. The
only other situation I can think of is that the two event-types
are not random with respect to the fixation trials (eg, remembered
items may be more likely to be followed by a number of fixation
trials, giving longer for the subject to process them), but I'm
not sure this would make much difference either in practice.

More importantly, I would check that your onset times are correct.
See previous helplist mailings about slice-timing, and reference
bins (fMRI_T0).

> And if we don't, should we say yes to
> the variable durations question?

No. Variable SOAs, yes; variable durations, no.


> 2)  high pass and low pass filtering:
> a)  the defaults for the high pass filter are different for different
> subjects and different sessions.  This is no doubt because of the
> different actual onsets (based on the different subsequent memory).
> But should we change them so that they are all the same?

If they differ little, it's unlikely to have much effect. I would
set them all the same if you intend to do a Random Effects
analysis (which assumes that each subject is analysed the
same way, which will not be precisely true in your case
with different designs/stimulus orderings, but you should
try to approximate). More importantly, they should not be
too much greater than 120s (a rule of thumb for the typical
cutoff for large increases low-frequency noise).


> b)  We have been using gaussian-4sec for low pass filtering.  When we
> redid our analysis without any low pass filtering we get much more
> robust activations for some contrasts and quite different activations
> for other contrasts.  What is going on here?  and what is appropriate
> for us to use?  If we use a 4 sec filter and our ITI is 3 seconds are
> we filtering out our signal rather than the noise?

If you don't lowpass filter (or use an AR(1) model) your statistics
will look much better, but are likely to be biased. This is because
you have not taken into account the temporal autocorrelation in the
data, meaning that your real df's are much smaller than the number
of scans (ie scans are not independent data points). For a multiscan
Fixed Effects analysis, you MUST perform some correction for
this autocorrelation. See numerous papers by Friston, Zarahn,
Bullmore and others.

Rik

--
---------------------------8-{)}-------------------------

DR R HENSON
Institute of Cognitive Neuroscience &
Wellcome Department of Cognitive Neurology
17 Queen Square
London, WC1N 3AR
England

EMAIL:  [log in to unmask]
URL:  http://www.fil.ion.ucl.ac.uk/~rhenson
TEL1  +44 (0)20 7679 1131
TEL2  +44 (0)20 7833 7472
FAX +44 (0)20 7813 1420

---------------------------------------------------------
--




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