[VoxBo] scaling adjusted trial-averages

[Daniel Y Kimberg]

Charan Ranganath wrote:
> Question: We've been getting adjusted time courses from VBview and
> because the data are mean-normalized, etc. there are negative and
> positive signal values in the time courses. This raises the question of
> how best to convert the data into % change or % change from zero. Any
> thoughts?

The two relevant operations are mean scaling (dividing each voxel time
series by its mean, separately for each tes file) and including an
intercept in the model.  When you do the former, your mean signal
value (and therefore your intercept) becomes 1 at every voxel, and the
observations in the mean-scaled signal are already proportions of the
mean signal (i.e., a time point with a value of 1.05 would have a
signal value 5% greater than the mean).  When you adjust for
covariates of no interest, and the intercept is included as a
covariate of no interest, the adjusted time series has the mean
subtracted off.  If you're already mean scaled, then it means you'll
see 0.05 for a time point that's 5% above the mean for that voxel.  If
not, then you can use the intercept to scale the adjusted values
appropriately.  Sometimes you can make life easier by designating the
intercept as being of interest, which wouldn't otherwise compromise
your model.

> Comment: On vbview, you can see fitted responses and covariate fits.
> However, the fitted response does not trial-average. The fits for single
> covariates do trial-average. Hwoever, if you shift-click to overlay
> these covariate fits with the trial-averaged time course, you can see
> that the two are on totally different scales. That's weird because in
> the old voxel surfer, the covariate fits and time courses appeared to be
> on the same scale, provided that you viewed in in terms of % change or
> raw signal.

I've fixed some but not all weird behaviors along these lines.  I'll
see if I can fix a few more before the upcoming release (which I keep
pushing back, but should just be another week or so).

dan