Dear Ged, Nick, Fatima,
scaling of the contrasts does not matter so both Ged's and my suggestions
are valid. The idea behind it is that scaled contrast vectors test the
same subspace of the design space.
Working it out is also the way I solved these issues, theory is explained
in detail in Ronald Christensen: "Plane Answers to Complex Questions"
(Springer texts in statistics).
On Fri, 14 Sep 2007, Ged Ridgway wrote:
> Ah, hello Volkmar,
> You managed to send this during the (long) time that it took me to write my
> waffley reply to Nick!
>> If you want to test differences between your two groups, you will have to
>> weight the subject factors as you would do for each group main effect. For
>> 2 groups with size N1 and N2 this would look like
>> 1/N1*ones(1,N1) -1/N2*ones(1,N2) 1 0 -1 0
> That's similar to what I'd guessed, except that I'd also "averaged" over the
> other factor, to give
> 1/N1*ones(1,N1) -1/N2*ones(1,N2) 1/2 1/2 -1/2 -1/2
> which of these do you think makes more sense? I hesitate to argue, but mine
> seems slightly more like the [1 1 -1 -1] that I think Nick was expecting...
> And do you have a good reference for this? I've mainly been trying to work it
> out as I go along ;-)
> All the best,
Department of Neurology [log in to unmask]
Universitaetsklinikum Freiburg Phone 49(0)761-270-5331
Breisacher Str. 64 Fax 49(0)761-270-5416
79106 Freiburg http://fbi.uniklinik-freiburg.de/