Dear Neil:
Thought I'd respond. There are of course many opinions about
smoothing both in space and time. Here is some of my understanding of
the issues though I suggest you look through the spm list and
published articles for more (erudite) info and others may want to add
their opinions.
Reasons for smoothing in space
1) Condition data to be more gaussian-like. This is definitely
necessary to use estimates from Gaussian Field theory for p-values.
The standard amount of smoothing is quoted to be 2x voxel resolution.
Without smoothing you should basically be using bonferroni correction
of your p values.
2) Reduce noise (based on matched filter theorem).
Both of these most likely improve detection of activations- but not
in every case, and there are examples in the literature of how using
various techniques of smoothing differently (scale space),
clustering, etc. may detect signals that spm missed. Although
smoothing will reduce the peak height of signals, by also improving
s/n ratio and the estimation of pvalues it usually helps. True it
does potentially reduce resolution, and may adversely affect the
detection of very focal signals.
Reasons for smoothing in time
1) In fmri the time points are certainly not independent. By
smoothing one introduces a known autocorrelation which allows proper
estimation of degrees of freedom.
2) Again probably reduces noise.
here too, oversmoothing could reduce transient signals, but on the
whole it probably improves detection.
Overall- if you want to detect focal signals in a small area of
brain, you may want to acquire high resolution data, not smooth it
and use the appropriate corrections for multiple comparisons (see
Zarahn/Aguirre/D'Esposito there are 2 articles from 1997 in
Neuroimage for some examples in addition to the usual Friston &
Worsley & Poline etc., etc. articles on these issues).
Hope this helps,
Darren
>I have the worry that by smoothing the data in an FMRI analysis,
>that if the activation that is present in the brain covers a v small
>area the smoothing process may wipe it out. Why would smoothing the
>data be a good idea - since there is an option in the stat tests to
>apply a threshold, and perform some sort of gaussian smoothing.
>
>Thank you
>
>Neil Smith
>City University
>______________________________________________________
>Get Your Private, Free Email at http://www.hotmail.com
Darren R. Gitelman, M.D.
-------------------------------------------------------------------------
Cognitive Neurology and Alzheimer's Disease Center
E-mail: [log in to unmask] WWW: http://www.brain.nwu.edu
Voice: (312) 908-9023 Fax: (312) 908-8789
Northwestern Univ., 320 E. Superior St., Searle 11-470, Chicago, IL 60611
-------------------------------------------------------------------------
|