[VoxBo] smoothed versus unsmoothed 1/f parameters

[Stacy Huffstetler]

Thanks, Dan.
For the new analyses, I was going to use vbfit to estimate the 1/f params
from each subject's power spectrum files (as generated by VoxBo).  If I'm
doing it that way, do I need to make sure to use the correct power
spectrums?  For instance, if I have smoothed and unsmoothed data,
thresholded and unthresholded data, and normalized and unnormalized data, do
I need to make a new 1/f params file for each combination?

Or is there a better way of doing it?

Stacy

On 11/16/06, Daniel Y Kimberg <kimberg at mail.med.upenn.edu> wrote:
>
> >Stacy Huffstetler wrote:
> > I was wondering what the difference was between the smoothed and
> > unsmoothed 1/f parameters.  I understand the theoretical difference,
> > that one was calculated using smoothed data and the other
> > unsmoothed, but I was wondering about the practical difference.  If,
> > for instance, I ran a whole bunch of GLMs with smoothed data, but I
> > included a 1/f noise curve calculated based on unsmoothed data
> > (e.g., unsmooth_params.ref), do I need to redo those analyses?  What
> > practical difference does it make?
>
> Two qualifications before I try to answer this question.  First, the
> 1/f params included with VoxBo are estimated from data taken from the
> old 1.5T GE scanner here at Penn.  You're probably a lot better off
> with an estimate derived from data that resembles what you're
> collecting a little more closely.
>
> Second, the reference of last resort on this is the set of two
> articles by Geoff and Eric in Neuroimage 5, pp. 179-212 ("Empirical
> analyses of BOLD fMRI statistics...").  Actually, Geoff himself would
> be the best source on this, except his spam filter apparently doesn't
> like this mailing list.
>
> All that said... eyeballing the prefab 1/f parameters suggests that
> you get a much steeper function (more power concentrated at the lowest
> frequency) with the estimates drawn from smoothed data, and that
> factoring out global signal has a smaller effect along those lines.
> So if I'm reading things right, when you use a 1/f draw from
> unsmoothed data even though you're looking at smoothed data, you're
> cheating yourself a little, because you're overestimating the amount
> of power at your task frequency(-ies) when H0 is true.  How big an
> effect this has... well, it seems like the cost to redoing the
> analysis is pretty small.  I would go ahead and re-run the analyses.
> But if possible, it would be even better to re-estimate the 1/f params
> from some more appropriate data.  Even if you don't have a run from
> this study that you can use, certainly someone in your building, or at
> least someone at Penn reading this list, will have some params from
> similar data.
>
> dan
>



-- 
Stacy Huffstetler
Research Assistant
University of Pennsylvania
Center for Cognitive Neuroscience
Dr. Matthew Botvinick's Lab
Phone:  215-898-2966
http://ccn.upenn.edu/~mmb
shuffste at mail.med.upenn.edu
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://www.voxbo.org/pipermail/voxbo-general/attachments/20061117/30b67220/attachment.html