I have some questions regarding design models for covariates in a
multisubject study.
We have 12 PET (H20 perfusion) scans per subject with a stress intervention.
Unfortunately the stressor is time dependent and can't be randomised so all
the subjects have an A A A B B B B B B A A A study protocol (or could be
A->L as time is of interest too).
One of our interests would be the correlation between perfusion and a
questionnaire or hormone response. Please correct me if I'm wrong: if I set
up a multisubject model with one condition & one covariate (eg 1 hormone
level) and set the covariate interaction *with subject*, I'm looking at
regions where subjects' perfusion correlates with their individual covariate
score.
1. If I set up a contrast with 1's over the covariate columns (same number
as subjects) then this gives me regions with significant positive
correlations - *but not for every subject* (I've checked). It makes sense
that the more subjects you include in the contrast the less strict the
correlation has to be within each subject to find significance (across all
the subjects or population). But if I'm not looking at significant
correlation for each & every subject I include in the contrast, what am I
looking at specifically. Is it, like ancova(?), treating the gradients as
the variable and looking for significant deviation from gradient=0?
2. The covariate may correlate with the protocol so I'm more interested in
the extra effects. I can use 2nd covariate (eg. 1 1 1 2 2 2 2 2 2 1 1 1...
[rpt]) to remove this effect. Should the interaction for this 2nd covariate
be set to "with subject" or "none"?
3. If the individual covariate shapes are the same with different magnitudes
of responses then it would be interesting to know if there are regions
across subjects that are associated. Should I set up the covariate with no
interaction to do this? Is this even a valid question?
Regards,
Joel
|