[VoxBo] 2 questions

[Daniel Y Kimberg]

Charan Ranganath wrote:

> (1) When using VBView to surf time course data, what is the default scale
> that is used? It looks like raw signal units, however, we get negative
> values in some averages, which makes no sense. However, this is seen after
> correcting for cov of no interest, so maybe that explains it. Anyway,
> please let me know if you know the scaling.

It is indeed raw signal units.  Typically when you remove the effects
of covariates of no interest, one of those covariates is the
intercept, so there will be lots of negative values, even during
trials.

> (2) [GEOFF, are you out there?] Here is a question that I posted earlier
> with no reply. I'm hoping to get a priming effect by repeating the
> question: I know that SPM deals with the whole autocorrelation thing with
> the AR(1) model. I believe (not sure though) that the model works by
> estimating the temporal autocorrelation of the residuals in a given GLM
> analysis. This would be useful for situations in which we don't have an
> empirical estimate from a separate dataset and don't want to use a stock
> function (estimated from the ancient GE scanners with prototype gradients
> from 1995). Here's my question, though: would we basically be doing the
> same thing if we fit a 1/f function to estimate autocorrelation from the
> data that we are trying to model? In other words, if you do a memory task
> and then put in the PS and reference function to fit a 1/f model to that,
> would that basically do the same thing as the AR(1) model?

I'll let others provide a more sophisticated answer, but here's my two
cents.  The problem with your scheme is that you're fitting your 1/f
model using a time series that's contaminated with your experimental
effect.  We'd expect this to make your test less sensitive (more Type
II error) by overestimating the amount of power at your experimental
frequencies when H0 is true.  This issue is mitigated a bit when your
experimental effect is small both in magnitude and extent (the 1/f
estimate is dominated by voxels where H0 is true).  And it's obviated
entirely for blocked designs, where a notch filter solves the problem.

I don't know enough about the AR(1) approach to say how SPM gets
around this.

I think a better solution would be to use the 1/f model would be to
find some blocked data from the same scanner and sequence, and
generate your own "stock" estimate.  I believe the recipe was posted
on this list at some point, perhaps someone with better access can dig
it up.

Bear in mind I'm a little disconnected from work right now, but I
don't think I'd give a less confused answer from my office.

dan