Alex,
> I have been following the ongoing discussion about the meaning of the
> beta values vis a vis signal change. I now have a pressing personal
> need to understand the beta values in a particular situation. I am
> doing a second level random effects simple regression of activation
> in a particular contrast with a behavioral measure. This yields a
> couple of clusters in the expected areas. If I click on the cluster
> of interest and plot the fitted and adjusted response versus an
> explanatory variable (my behavioral measure) I get a linear
> regression plot with the behavioral score on the X-axis and "Effect
> Size" on the Y-axis. These values range from about -8 to 10 (across
> the 10 subjects). I do not think this can represent % signal change!
> Also, interestingly, running a completely different regressor against
> the same data yields Y-axis values in the -1 to 1 range. (Is this
> related to the fact that the second regression was less statistically
> significant?)
>
> I have previously submitted this data for publication with the Y-axis
> labelled "effect size" and was critiqued by one reviewer for the
> scale being "arbitrary". Any suggestions on what this means and how
> to most accurately label the data would be greatly appreciated.
The parameter estimates (at the first or second level) are simply
regression coefficients. The reason they can, in some instances be
interpreted as percent whole-brain activations is that the regressors
contructed by SPM are scaled so that a paricular state or event
produces a unit response when the parameter estimate is one. If your
regressor is not constrained like this (e.g. entering raw reaction time
data) then the parameter estimates have to be interpreted in relation
ot he scaling of the covariate: They are simply the amount of
activation (in percent whole brain signal) induced by a unit change in
the covariate. If you used milliseconds to express the reaction times
the parameter estimates will be a thousand times smaller than if you
had used seconds.
I hope this helps - Karl
|