Greig -
F-constrasts of the type [eye(N) -eye(N)], to compare
two event-types modelled with N FIR bins, are appropriate
for detecting any difference between the two event-related
responses. However, there is nothing wrong with testing
more focused contrasts if you want to make more constrained
assumptions about the shape of the response: for example,
that it peaks between 2-8s poststimulus (when allowing for
slice-timing differences of 4s, as in the Scandinavian Journal
of Psychology article you mention), and hence use t-contrasts
averaged over a subset of bins. You could even constrain
your constrasts to look for "canonical" responses by choosing
a "canonical-weighted" constrast of FIR parameters, eg:
[spm_hrf(b) -spm_hrf(b)], where b is the FIR bin-width and
spm_hrf.m is the SPM function that returns the canonical HRF
(assuming in this case that N*b=32s, the default duration
of the canonical HRF). These more focused constrasts will
generally buy you more power at the expense of missing
other shapes of event-locked responses.
The problem comes when you want to do second-level tests
(eg treating subjects as "random effects"). Though you can
put the N FIR parameter estimates into a 2XN repeated-measures
ANOVA (using the SPM PET interface), and test for an event-type-
by-time interaction, you will need to make some correction for
nonsphericity (the fact that the N parameters are unlikely to be
independent or identically distributed over subjects - this is why
Andrew labelled such designs "'Dodgy' population inference"!).
We have been working on this in the development version of SPM
(see Glaser et al, HBM 2001), in the context of more general
variance-component estimation, but it is not quite ready for release.
Alternatively, you could use the "hidden" SPM feature of multivariate,
second-level statistics, which evade (or "eschew" perhaps? ;-)
the problem of nonsphericity - see our HBM2000 abstract on my
webpage if interested - though are generally less powerful.
Note one nice thing about the "canonical-weighted" contrasts on
an FIR model that I mentioned above is that you can take the
resulting con*imgs into a second-level, one-sample t-test, just
as if you can used an canonical HRF basis function in the first-
level model, rather than the FIR basis set. Provided that the binsize
equals the effective peristimulus sampling rate, you will get
mathematically the same parameter estimate for the canonical-
HRF-first-level-model as you do for the canonical-HRF-weighted-
constrast-on-an-FIR-first-level-model. The nice thing about putting
constraints on the shape of the HRF into the contrasts of an FIR
model, rather than into the original regressors (at least for
subsequent
second-level analyses), is that you can test for different shaped
responses simply by adding new contrasts, without needing to
re-fit a (typically large) first-level model.
Hope this makes sense
Rik
Dear Rik,
I note that your HBM abstract describes F-contrasts with the FIR
method and
that the SPM99 manual also suggests F-contrasts are need for
meaningful
inference (p. 39). However, your paper on repetition priming in the
Scandinavian Journal of Psychology reports t-contrasts performed on
parameter estimates from several bins. This suggests that a
second-level
analysis might be conducted with t-contrast images derived using
the FIR
method. Could you clarify whether this is the case?
Regards,
Greig
> From: Rik Henson <[log in to unmask]>
> Reply-To: Rik Henson <[log in to unmask]>
> Date: Thu, 6 Sep 2001 18:22:50 +0100
> To: [log in to unmask]
> Subject: Re: FIR basis?
>
>
> Attached is an extended version of spm_get_bf.m that
> includes an FIR basis set for event-related analyses (which
> prompts for binsize, in seconds, and number of bins - see
> SPM manual for more details) and that should work fine
> with the released version of SPM99.
>
> For the pros and cons of an FIR model to make inferences,
> see Henson et al. (2001), HBM01 abstract, Neuroimage, 13, 149,
> though note erroneous labelling of figures in published version;
> correct version available from:
>
> http://www.fil.ion.ucl.ac.uk/~rhenson/refs.html
>
> and for the use of an FIR model to plot data as a function of
> peristimulus time, see recent posting by Russ Poldrack:
>
>
http://www.jiscmail.ac.uk/cgi-bin/wa.exe?A2=ind0108&L=spm&P=R6690
>
> http://www.nmr.mgh.harvard.edu/~poldrack/spm/pst_avg.html
>
> Best wishes
> Rik
>
--
Dr Greig de Zubicaray
Centre for Magnetic Resonance
The University of Queensland
Brisbane, QLD 4072
AUSTRALIA
Tel: +61 (0) 7 3365 4250 [direct]
+61 (0) 7 3365 4100 [CMR]
Fax: +61 (0) 7 3365 3833
--
---------------------------8-{)}-------------------------
DR R HENSON
Wellcome Department of Cognitive Neurology
& Institute of Cognitive Neuroscience
University College London
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
MOB +44 (0)794 1377 345
FAX +44 (0)20 7813 1420
---------------------------------------------------------
--
