Dear Rachel,
On 2005-05-04 (Wed) at 11:50:34 +0100, Rachel Mitchell <[log in to unmask]> wrote:
> Hi
>
>
> Having generated individual con images for each participant for each of the
> 3 tasks I then proceeded to the within-subjects ANOVA option
> I entered the 3 comparison a con images (task 1, task 2, task 3) for subject
> 1, then subject 2 and so on
> I wasn't 100% convinced that I might have broken the sphericity assumption,
> but thought I'd go along with it anyway and selected the correction - was
> this wrong?
> By the phrasing of the next question I was a little confused as to whether I
> should select repetitions across subjects or conditions. I selected subjects
> (but to be honest it wouldn't work even when I tried it across conditions) -
> was this wrong?
No, you chose the correct option. "Replications" referes to independent
repliation of the experiment, or the random effect.
> I said yes to the following correlation condition (from the list, if you
> have selected a sphericity correction most people seem to be saying to go
> ahead and automatically say yes to this) - was this wrong?
Again you chose the correct option, since one can expect that the
different conditions in a single subject are correlated.
> Parameter estimability was greayed out for all variables but I gather that
> isn't actually a problem.
>
> Now at the contrast manager stage, if I had run a between subjects ANOVA, I
> used to enter '1 0 0' to determine the effects of task 1, '0 1 0' for task
> 2, and '0 0 1' for task 3
> I'd then select all 3 together and hey presto i got my conjunction
> In a within-subjects ANOVA when I entering '1 0 0' for task 1, I get and
> invalid contrast error message
> So have I made a mistake here too?
If you want to detect a main effect in a within-subject ANOVA you alos
have to put weight on the subject constants otherwise the contrast is
unestimable. Intuitively, you can think about it like this: your 3 task
regressors only model the deviations from the subject mean, which itself
is modelled in with the subject constants. Thus, in order to calculate
a main effect you have to have the subjects' mean plus the
condition-specific deviations from it.
Computationally, the sum of weights on the subject constants have be
equal to the sum of the weights on the condition regressors. Thus, in
your design (3 task regressors, 15 subject constants) your contrast for
the main effect of task 1 is: [1 0 0 (1/15)*ones(1,15)]
Note, that this also holds true for "negative" main effect. Namely,
[-1 0 0 -(1/15)*ones(1,15)]
The statistical logic of the contrasts specification is detailed and
exemplified in the case of a t-test in the following document:
http://www.pallier.org/ressources/glm_anova/glm_anova2.pdf
Cheers,
Jan
>
> Thank-you for your valuable time
>
> Regards
> Rachel Mitchell
>
>
> -----------------------------------------------------------------------
> Dr Rachel L. C. Mitchell
> Lecturer in Cognitive Psychology, University of Reading
> Honorary Research Fellow, Institute of Psychiatry
> Research Psychologist, Berkshire Healthcare NHS Trust
>
> Correspondence Address:
> School of Psychology
> Whiteknights Road
> University of Reading
> Reading
> Berkshire
> RG6 6AL
>
> Tel: +44 (0)118 378 8523
> Direct Dial: +44 (0)118 378 7530
> Fax: +44 (0)118 378 6715
> -----------------------------------------------------------------------
--
Jan Gläscher NeuroImage Nord
+49-40-42803-7890 (office) Institute for Systems Neuroscience, Bldg S10
+49-40-42803-9955 (fax) University Medical Center Hamburg-Eppendorf
[log in to unmask] Martinistr. 52
20246 Hamburg
Germany
http://www.uke.uni-hamburg.de/kliniken/neurologie/index_16969.php
----------------------------------------------------------------------------
GnuPG/PGP key id: FEC4B55C
fingerprint: 5A36 1EF6 8472 117E 805A F240 3146 A410 FEC4 B55C
----------------------------------------------------------------------------
|