Hi Anson,
Some of Tom Nichols' papers I found helpful are listed with links here:
http://brainvis.wustl.edu/pipermail/neuro-mult-comp/2005-December/000001.html
Donna
On 02/06/2006 03:06 AM, Marko Wilke wrote:
> Hi Anson,
>
>> May I know what's the different between FDR, FWE and none when SPM2
>> asks about 'p value adjustment to control'?
>
>
> Wow, that's a big one :) Basically, this is the famous "correction for
> multiple comparisons" issue. When you do multiple statistical tests,
> the more you do, the more likely your are to get false positive
> results just by chance. Therefore, you need to correct for this
> so-called Type 1 error, either using the family-wise error (FWE) or
> the false discovery rate (FDR) approach. Uncorrected does not do
> anything here to correct for these multiple tests and should thus only
> be used to explore results. You will find hundreds of hits on the
> mailbase (1) if you search for this.
>
>> The 3 attachments are generated using FDR, FWE and none respectively.
>> The results are quite different. The scan was a blank EPI scan that
>> the subject was asked to relax and laying still with eye closed. So I
>> didnt expect any activations seen in the glass brains.
>
>
> Well, REST is only random episodic silent thinking, the brain is not
> really swithced off :) Also, from your desing matrix, it does look
> like you do have multiple conditions when you enter 1 as a contrast,
> so check that you really only look at this condition.
>
>> Which one (FDR, FWE or none) should I use?
>
>
> None is not good enough, really. Whether FWE or FDR is a matter of
> discussion: the former is more specific, the latter more sensitive.
> Both are statistically legitimate. You may want to read some of Tom
> Nichols papers on this issue, which even I (and that says a lot!)
> understood.
>
> Hope this helps,
> Marko
>
> (1) www.jiscmail.ac.uk/cgi-bin/webadmin?S1=spm
|