Dear Mike (& Karl),
At 13:43 01/08/99 +0100, Karl Friston wrote:
| At 15:50 30/07/99 +0100, Dr Mike Glabus wrote:
| > In pre-SPM99 versions, if one wanted to do an interaction analysis of
| > e.g. a neuropsych. score measured at (scan) condition 1 and condition
| > 2 , the approach was to construct an interaction matrix by subtracting
| > the scores at condition 1 from condition 2 then constructing an
| > interaction covariate thus:
| >
| > [+diffscoresubj1, - diffscoresubj1,...
| > +difscoresubj n, -diffscoresubj n]
| >
| > and then use this in a multi-subject, conditions & covariate design as
| > the covariate of interest.
| >
| > However, in the PET/SPECT toolbox of SPM99, there is now an option to
| > do an interaction analysis with the covariate of interest. How is this
| > invoked? I've tried entering the actual clinical scores for each
| > condition in the single matrix requested, rather than the +diff and -
| > diff as before. Is this correct? Also, I note that after I request an
| > "interaction by condition" analysis, the printed design for the
| > General Linear Model now displays two separate columns for the single
| > covariate labeled 'covar'@condition1 and [log in to unmask] Will this
| > design invoke the same interaction analysis as by the 'old' method?
|
| You should wait for confirmation from Andew but I think the PET
| inferface allows for interactions by modeling each level of one factor
| in a separable fashion under each level of the other (e.g. condition
| or subject). This is exactly the same statistical model as you might
| have constructed before but the interaction effect is now tested
| explicitly with a contrast (in your case) 1 -1, testing for the
| difference in terms of one factor under the different levels of the
| other. Notice that this model implicitly also models the main effect
| of the covariate (tested for with 1 1).
This is correct. As Karl says, both models are equivalent.
The advantage of the SPM99 (PET/SPECT) way of modelling (also available in
basic Stats options) is that all the covariate centering and stuff is
handled for you: By default (for covariate by condition modelling for
example) covariates are centred within condition (factor), so that the
covariate columns only model effects within condition, and therefore don't
"steal" some of the main effect of factor.
Also, factor by covariate interactions for factors with more than two
levels are easy to do.
Hope this helps,
-andrew
+- Dr Andrew Holmes ------------------ mailto:[log in to unmask] -+
| ___ __ ___ Robertson Centre for Biostatistics |
| ( ,) / _)( ,) Boyd Orr Building, |
| ) \( (_ ) ,\ University Avenue, |
| (_)\_)\__)(___/ Glasgow. G12 8QQ Scotland, UK. |
+----------------------------------------- http://www.rcb.gla.ac.uk/ -+
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|