Torsten, not sure if this is related to your question, but...
Note that the regression coefficient (i.e. beta value) from GLM and Pearson's correlation coefficient (aka r, coefficient of determination) tell you two very different things. Pearson's correlation coefficient tells you how linear the relationship between two variables is, but it tells you nothing about the magnitude of the relationship (i.e. slope of the best fit line). The regression coefficient is the slope of the best fit line (and is readily extended to the multivariate case), but it tells you little about how linear the relationship is or how much your data deviate from the best fit line. Scale the regression coefficient by the standard error and you get a t-value (or effectively a z-value for large n) that tells you someting about the slope *and* how well it fits your data.
I have seen r values transformed to z values (a la Fischer's z-transform) used more often for seed-based voxel correlation than I have seen the regression coefficients used. My personal view is that your choice should be motivated by your hypothesis and how sensitive each measure is to noise -- not just what others have used in the past.
If you are interested in getting a multivariate r value, check out these publications:
Nagelkerke 1991: A Note on a General Definition of the Coefficient of Determination
Magee 1990: R^2 measures based on Wald and likelihood ratio joint significance tests
On Thursday, September 13, 2012 07:34:00 you wrote:
> Dear Torsten,
>
> The GLM outputs, even when used with normalised inputs or divided by the standard error, are not the same as correlation values. They are similar, and statistical tests are identical, but if you need an r value to convert to Z with the Fisher transform, then you need to calculate it using the full formula for correlation which is subtly different. Or you can just use Z values output by the GLM directly.
>
> All the best,
> Mark
>
>
>
> On 6 Sep 2012, at 06:32, Torsten Ruest <[log in to unmask]> wrote:
>
> > Dear Steve,
> >
> > thanks for your comment - I thought maybe it's just some rounding errors . I've tried to find the error that may explain the r value higher than 1, but...
> >
> > I used the method described in the mailing list as well as elsewhere:
> >
> > 1. regress nuisance variables
> > 2. take the residual (1.), add back the mean, normalise the latter by dividing by it's standard deviation, then extract the seed mean time timeseries.
> > 3. add the normalised "meaned" residual and the normalised "meaned" timeseries for that seed to the feat design and go.
> >
> > I also tried the other way I read here:
> > 1. add normalised "meaned" timeseries to feat
> > 2. add unnormalised "meaned" residual to feat
> > 3. divide resulting cope / varcope by standard deviation
> >
> > the results are virtually identical.
> >
> > I take it that after normalization of the residuals and timeseries, the glm output should be in the range -1:1, so for isolating the error, whatever happened before, eg during nuisance regression etc, is kind of irrelevant ?
> >
> > Thanks in advance.
> >
> > Best wishes,
> >
> > Torsten
> >
>
|