Hi,
I think it depends if you are more "worried" by the false-positives or
by the false-negatives.
FWE has got less false-positives than FDR but more false negatives.
For a publication I think it would be better to use FWE, so that you are
more certain that what you see is real.
I would feel uncomfortable reading a paper with different corrections
for different groups. It would seem to me like comparing pears with
apples. I am not sure, but maybe it would be all right if you present
the same correction and then for one group also a different approach,
explaining why you do it? I am not too sure about the different
p-levels.
In any case, can't you do a 2-sample t-test and compare controls vs
patients and present the results of the comparison?
Regards,
Laura
-----Original Message-----
From: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]]
On Behalf Of Niklas Lenfeldt
Sent: 05 April 2007 11:42
To: [log in to unmask]
Subject: [SPM] FWE or FDR? few or many voxels?
Hi everybody
I have a question on which statistics to use. We have used a
finger-tapping task to model cerebral activation in patients and
controls, however, there are 50% more patients than controls. Using
FDR=0.05 yields activation in the region of thousands of voxels
especially for the patients. I wonder, should I increase to FDR=0.005 or
even more or should I use FWE instead? I am intrested in displaying the
local peaks so I do not want oceans of voxels, especially since, as I
understand it, differences in cluster size could be attributed to the
differences in group size. Do FWE and FDR account for group size in the
same manner? Would it be alright to present data using different
p-levels for patients and controls, as long as I focus on peaks and not
on cluster size? Could I use the same p-levels but using FWE for one
group and FDR for the other? What is to prefer, strong statistical
levels giving smaller clusters and displaying all clusters or looser
stat levels giving larger clusters and thresholding instead?
If someone has an idea on this I would be most greatful, including
telling me where to find more information on the issue.
Regards,
Nick
**********************************************************************
This email is confidential and is intended solely for the person or entity to whom it is addressed. If this is not you, please forward the message to [log in to unmask] We have scanned this email before sending it, but cannot guarantee that malicious software is absent and we shall carry no liability in this regard.
We advise that information intended to be kept confidential should not be sent by email. We also advise that health concerns should be discussed with a medical professional in person or by telephone. NHS Direct can also provide advice. We shall not be liable for any failure to follow this advice. University College London Hospitals NHS Foundation Trust (UCLH).
|