"Chris Gottschalk, MD" wrote:
> I am trying to define contrasts for the first stage of a RanFx analysis of
> SPECT data-- two conditions per subject, one scan per condition. I have 75
> subjects, distributed across treatment conditions. I would like to examine
> the groupXcondition interactions in an appropriate way.
>
> I enter these data into the "Multi-Subject: subj X condition interaction &
> covariates" model-- planning to ignore the condition effect until the
> second level. However, I can't get past the first level: with this model, I
> consistently get a GUI error.
>
> Is this a software problem on my system, or a design problem?
>
> I am concerned that this model would produce as many columns as scans, and
> so be inestimable or "over-determined"?
I think you need to formulate the analysis in a different way.
In SPM, choose the 'Multi-group: conditions & covariates'
design. By choosing a suitable contrast function you will
immediately be able to examine the
interaction between group (treatment condition)
differences and condition differences.
You do not need a 'second-level' of analysis.
A 'two-level' analysis would be appropriate if you
had more than one scan per condition per subject. In the
first level you could then estimate the variance
of the difference images within each
subject (that is, the difference between conditions images),
and in the second-level estimate the variance between-subjects.
This is called a random-effects analysis, because the
images from the first stage are treated as random variables in
the second stage (we can do this as we know their mean and variance).
But because you have only one scan per condition, this isn't the
appropriate analysis.
If you'd like to read more around this topic see
http://www.mailbase.ac.uk/lists/spm/1999-06/0097.html
and references therein.
So, to re-cap, use the 'Multi-group: conditions & covariates'
design and do everything in one stage.
Yours,
Will.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|