Print

Print


Dear Kim,

That's right, T=Inf is the correct threshold.  I.e., there is *no* threshold that will control FDR at 5%.  

This might help:

load xSPM
Ps=xSPM.Ps;
S=xSPM.S;
P_FDR = spm_P_FDR(spm_invNcdf(1-Ps(1)),[1 Inf],'Z',1,Ps')
subplot(2,1,1)
plot((1:S)/S,Ps,'-o',(1:S)/S,(1:S)/S*P_FDR);xlim([0 0.5]);ylim([0 0.05])   
title('Observed P versus Expected')      
subplot(2,1,2)
loglog((1:S)/S,Ps,'-o',(1:S)/S,(1:S)/S*P_FDR);xlim([0 0.5]);ylim([0 0.05])  
title('Observed P versus Expected, log-log')      
       
As you'll see, P_FDR is 0.1789.  If you ask for FDR control at 0.2, you should see some voxels.

-Tom


On Thu, Jul 21, 2011 at 12:38 PM, Guillaume Flandin <[log in to unmask]> wrote:
Dear Kim,

I cannot see anything wrong, you seem to be in the situation where there
is no signal in your data and the BH procedure, as implemented in SPM,
returns Inf in this case.

spm_uc_FDR(0.05,xSPM.df,xSPM.STAT,xSPM.n,xSPM.Ps')

See slide 43 of this talk:
http://www.fil.ion.ucl.ac.uk/spm/course/slides11/06_RFT_etc_FIL2011May.ppt

You can also compute:
 spm_invTcdf(1-0.05/33012,24) = 6.0479
which is the 0.05 critical threshold, that your data do not reach.

I copy this email to Tom Nichols who might be able to tell us more.

Best regards,
Guillaume.


On 21/07/11 11:55, Veroude, K. wrote:
> Dear Guillaume,
>
> Yes I changed topoFDR to 0 in SPM_defaults. Attached you can find the xSPM.mat.
>
> Thanks,
>
> Kim
>
> ________________________________________
> Van: Guillaume Flandin [[log in to unmask]]
> Verzonden: donderdag 21 juli 2011 12:12
> Aan: Veroude, K.
> Onderwerp: Re: [SPM] FDRp: Inf in SPM 8
>
> Dear Kim,
>
> so I assume that you are using voxel-wise FDR (ie you switched the
> topoFDR option to false)?
> When you are looking at this result, there should be a variable called
> 'xSPM' in the MATLAB workspace, could you send it to me?
>>> save xSPM.mat xSPM
>
> Many thanks,
> Guillaume.
>
>
> On 21/07/11 08:32, Veroude, K. wrote:
>> Dear Guillaume,
>>
>> This is the Results window we get when applying FDR correction at p < 0.05. Any help would be appreciated.
>>
>> Best wishes,
>>
>> Kim
>>
>>
>> -----Oorspronkelijk bericht-----
>> Van: Guillaume Flandin [mailto:[log in to unmask]]
>> Verzonden: woensdag 20 juli 2011 16:27
>> Aan: Veroude, K.
>> Onderwerp: Re: [SPM] FDRp: Inf in SPM 8
>>
>> Dear Kim,
>>
>> can you give me more details on what you are observing? A screenshot of
>> the Results Graphics window would be useful to start with.
>>
>> Best regards,
>> Guillaume.
>>
>>
>> On 20/07/11 09:40, Kim Veroude wrote:
>>> Hi,
>>>
>>> We have a similar problem, when applying FDR correction in SPM8 it
>>> shows T = Inf. Have you found a solution for this?
>>>
>>> Thanks!
>>>
>>> Kim
>>>
>>>
>>
>>
>
>
> --
> Guillaume Flandin, PhD
> Wellcome Trust Centre for Neuroimaging
> University College London
> 12 Queen Square
> London WC1N 3BG


--
Guillaume Flandin, PhD
Wellcome Trust Centre for Neuroimaging
University College London
12 Queen Square
London WC1N 3BG



--
__________________________________________________________
Thomas Nichols, PhD
Principal Research Fellow, Head of Neuroimaging Statistics
Department of Statistics & Warwick Manufacturing Group
University of Warwick, Coventry  CV4 7AL, United Kingdom

Web: http://go.warwick.ac.uk/tenichols
Email: [log in to unmask]
Phone, Stats: +44 24761 51086, WMG: +44 24761 50752
Fax:  +44 24 7652 4532

Now out!  
Handbook of fMRI Data Analysis, by RA Poldrack, JA Mumford & TE Nichols
http://go.warwick.ac.uk/tenichols/publications/books




--
__________________________________________________________
Thomas Nichols, PhD
Principal Research Fellow, Head of Neuroimaging Statistics
Department of Statistics & Warwick Manufacturing Group
University of Warwick, Coventry  CV4 7AL, United Kingdom

Web: http://go.warwick.ac.uk/tenichols
Email: [log in to unmask]
Phone, Stats: +44 24761 51086, WMG: +44 24761 50752
Fax:  +44 24 7652 4532

Now out!  
Handbook of fMRI Data Analysis, by RA Poldrack, JA Mumford & TE Nichols
http://go.warwick.ac.uk/tenichols/publications/books