Dear Didine,
> Dear SPM ers:
>
> I am trying to find changes in binding of a PET ligand from baseline
> to activation (paired studies) where a covariate is obtained during
> the "activation", but the baseline covariate is zero.
> I would like to correct for the effect of intersubject differences in
> covariate on binding during activation and look at correlations
> between differences in binding and the covariate.
>
> In SPM99, I used multisubject: conditions and covariates and entered
> covariates as follows:
> first mean center covariates for activation which results in : a b c .....n
> second enter covariate as follows: 0 a 0 b 0 c .....0 n
> contrasts are [1 -1 0] & [-1 1 0] for condition effect (at mean
> covariate value) after adjusting for covariate
> and [0 0 1] & [0 0 -1] for positive and negative correlation.
>
This all looks very good. Your first contrasts check for condition effects, after
any variance attributable to your covariate has been removed, and your second
pair of contrasts looks for correlation between binding and the covariate.
>
> however I had 3 options for global norm:
> ANCOVA BY SUBJECT, PROPORTIONAL SCALING and NO NORMALISATION.
> ANCOVA BY SUBJECT cannot be modeled because there is only one scan
> per condition.
>
The "ANCOVA by subject" will devour one degree of freedom for each subject.
Between this and the subject effects there wont be much left, hence your limited
choice.
>
> So the only options remaining are PROPORTIONAL SCALING and NO NORMALISATION
>
> Based on SPM99 choices it seems that ANCOVA is inappropriate for the
> design (paired with or without a covariate) as above with subject
> having .
> Can someone explain why?
>
I am not 100% I understand the question. Are we talking ANCOVA as a means of
removing global variance? I will assume we are.
You can think of it like this. If we consider a given voxel, then for each
subject we have two points. Think of these two points as two points in a
scatter-plot with global activity on the x-axis, and voxel (local) activity on
the y-axis. What we do by including the "subject effects" is to fit a mean to
these points. The ANCOVA-by-subject covariate on the other hand will fit a
straight line between these points. Hence, with these two "effects" we have
modelled the data perfectly (i.e. we have fitted a straight line to two points).
Now what the statistics aim at elucidating is if the model is a good description
of the data. Fitting a straight line to two points will always "look good", but
there is really no way of determining if it was a "good" model or not.
You have three options;
1. Go for proportional scaling.
2. Skip global normalisation. When doing PET ligand stuff statistical analysis
has often been preceeded by some modelling step where raw counts have been
transformed to a measure like e.g. binding potential. It could well be argued
that this "quantification" step obliterates the need for global normalisation.
This would also have the affect that you are able to look for "actual" changes
rather than just "relative" (what areas change with our condition in a way that
is different from that of the majority of the brain).
3. Create "your own" global regressor and go for good old fashioned ANCOVA (i.e.
not by subject). You do this by picking the SPMcfg.mat file from one of your
previous analyses. In there you will find the "global activity" of all your scans
in xGX.rg. Enter this as a confound.
I hope I understood you question properly.
Good luck Jesper
|