Leighton,
> Quick follow-up: Centering both covariates to the overall mean
> should de-mean them - right?
Yes, precisely
> The SD/mean I get still seem (about 4 times) too big. Has anyone else
> done this and got reasonable results?
Don't know... where do you have experience from this? From ROIs?
-Tom
-- Thomas Nichols -------------------- Department of Biostatistics
http://www.sph.umich.edu/~nichols University of Michigan
[log in to unmask] 1420 Washington Heights
-------------------------------------- Ann Arbor, MI 48109-2029
> > > I would like to characterise the intersubject noise (variance after
> > > adjusting out effects) in a population of SPECT scans. I
> > > have computed the coefficient of variation (SD/mean) image using
> > > SQRT(ResMS)/beta_0003 (there were 2 covariates in the study).
> > > Is this a valid approach?
> >
> > It depends on the interpretation of beta_0003, the intercept. If all
> > of the covariates are mean zero, then beta_0003 is precisely the mean
> > of the SPECT data at each voxel. If the covariates aren't de-meaned,
> > then beta_0003 is... the intercept, the fitted response when all
> > covarariates are zero.
> >
> >
> > > The numbers in grey matter seem a bit big - around 0.6
> > > Are the units for the values in SQRT(ResMS) and beta_003
> > > the same or is some kind of scale factor involved?
> >
> > When the covariates are de-meaned it should be exactly the coefficient
> > of variation, no scaling needed.
> >
> > -Tom
|