Dear Stuart,
> Dear All:
>
> I am trying to implement a second level analysis following a first
> level covariate analysis using stimulus rating for each scan as my
> covariate of interest. The model is, thus - single subject:
> covariates only. To put this another way, I have a covariate design
> that is unbalanced, how do I proceed to the second level?
>
If I understand you right you have a number of scans performed under some
condition where subjects are asked to rate some sensation (I don't want to
know which!) during scanning. Hence, for each subject you have a set of
ratings ranging e.g. from 1 to 10, and you enter these into a single
subject-covariates only model (I guess you could equally well use a multi
subject-covariates only design with a covariate-by-subject interaction)
yielding for each subject a map of the dependence between rCBF and the
stimulus rating. To take this to a second level analysis you simply create
the *con images for each subject (in this case identical to the beta images)
and enter these into a one-sample t-test.
I am not sure what you mean by an unbalanced design. Do you have unequal no.
of scans for different subjects, or do you have a very poor range of ratings
for some subjects? If so your estimates from the first level analysis will
be of varible quality (i.e. be assosciated with different estimation
uncertainty) which is in principle a violation of the assumption behind the
two-step implementation of the RFX model. This is similar to the situation
with retrospective assignment of events in ER fMRI designs, where the
proportion of different event types can sometimes be quite variable between
different subjects. I think the concensus there is that as long as that
difference is within reasonable bounds the ensuing violation is of little
consequence.
I guess you could make a similar judgement here, i.e. discarding subjects
with very few scans compared to the rest, and also subjects that e.g. rate
all scans as a 5, keep the rest and cross your fingers.
>
> In addition I have succesfully completed a second level analysis
> using a straight contrast of two conditions and then plugging the
> con_...?? images into the basic model single group t-test. All looks
> great and I am pleased that my signal has not vanished. I am now
> trying to use the two-sample t-test option to compare two groups but
> I never have any parameter space left... I enter the con_..?? images
> from both groups, assign the two groups and am then left with no
> parameter estimability.
>
This sounds very strange. As long as your group sizes are larger than one
subject it should work. Perhaps you could describe exactly what you do
(perhaps send the SPMcfg.mat file).
Good luck Jesper
|