Dear Matthew,
May I ask a follow-up question on this topic ..
How do you get the FWE corrected threshold for a combination of
voxel-level threshold and a given cluster size other than performing a
small volume correction in SPM, or do you refer to this here ?
Thanks,
Roland
Roland Zahn, MD PhD
Research Fellow
NIH / NINDS
Cognitive Neuroscience Section
Building 10, MSC 1440
Room 7D43 (Postal address)
Room 7D49 (Physical address)
Bethesda, MD 20892-1440
Tel.: (+1)-301-402-6392
Fax.: (+1)-301-480-2909
> -----Original Message-----
> From: Matthew Brett [mailto:[log in to unmask]]
> Sent: Monday, February 26, 2007 4:49 PM
> To: [log in to unmask]
> Subject: Re: [SPM] Correct thresholding of results: when is p < 0.001
> acceptable?
>
> Hi,
>
> On 2/26/07, Douglas Burman <[log in to unmask]> wrote:
> > Brett (and any statistician who would like to add their thoughts):
> >
> > I wonder if you could comment about the use of uncorrected
thresholds in
> > combination with extent thresholds.
>
> Well, it's really the same argument; obviously the extent threshold
> gives you more specificity to larger activations, but it remains
> unquantified in a statistical sense. You can calculate the combined
> estimated FWE for a given uncorrected height / extent threshold with
> the SPM machinery though.
>
> >Another lab on campus has run
> > monte carlo simulations on various intensity / extent combinations
> (using
> > AFNI), and found a false discovery rate < 0.05 whenever the
uncorrected
> > Z-statistic is greater than about 4.00
>
> I can't comment on the AFNI Monte-Carlo procedure, I'm afraid I don't
> know anything about it. The equivalent in SPM would be the
> permutation test method of correction in SnPM - my own personal
> favorite. The problem is that if you apply a threshold calculated from
> one dataset to another, you need to be absolutely sure that the noise,
> smoothing, degrees for freedom, underlying data smoothness, field of
> view, etc are the same, otherwise that threshold will not be valid.
>
> I think the principle would be - if you want to do an FWE correction,
> use the FWE machinery - again I would recommend SnPM strongly for
> random effects. If you want a more informal threshold, then justify
> that in the paper with suitable argument, and I would hope your
> reviewer would take that seriously...!
>
> Best,
>
> Matthew
|