Dear Chaorui,
>Thank you very much for your answer. It is quite clear. But I have another
>minor question that I did not get answer from other people previously.
>For this PET study, we would also like to know how age affect the change of
>brain metabolism. However, when we use age as a covariate, the program
>always give error message. We suppose that is because the change of age is
>highly correlated with the change of brain metabolism overtime. So, in that
>case, we should not put 'age'as a covariate in the model. Then, we decided
>to look at how the initial age (the age at the first investigation) can
>affect the longitudinal metabolic change during these 6 years. In other
>words, we would like to see if the 'initial age' confounds our findings.
>The problem we have right now is that we do not know how to fulfill this
>idea in SPM analysis. Because SPM requires one covariate for each scan. But
>in our study, we have 3 scans for each subject(or 3 conditions, because we
>have 3 time points) and one covariate (the initial age). We do not know how
>to enter 'age' as a covariate in the model. We would like to know if SPM
>can handle this kind of problem.
Yes, there should be no problem with entering age as a confound or nuisance
variable. I am not sure why you obtained a warning with the first model
you tried.
It may be because you used to same age for each subject and modelled subject x
age interactions. These effects are exactly the same as the subject
effects and
are consequently inestimable. Age could enter as a main effect, or through an
interaction with time effects. To model the first, simply include
an extra regressor with the initial age of each subject (but not modelling
interactions of age with subject). To model the second effect, simply
multiply
the age regressor with the main effect of time. For example if the data
were entered as subj1-scan1 subj1-scan2,....
[age1 age1 age1 age2 age2....].*[-1 0 1 -1 0 ...]
This would model the interaction age x linear time effects (i.e. how
time-dependent reduction in metabolism depended on age.
You could model the age x time x subject effects, but this may result in
a large, less sensitive mode.
I hope this helps - Karl
|