Dear Guillaume,
Thank you very much for your response and for confirming that results are valid.
I had a look at the slides, they spoke to me as I also read Genovese & al. 2002.
From what I understand, p-values are sorted and those being below a threshold calculated from P <= i/V * q/c(V) are declared as "active".
However, it remains unclear to me how those "validated as active sorted p-values" are being assigned a FDR p-value (i.e. how p-values are assigned to green and red dots), and why SPM chooses the minimum value in (min(QS(I:S))?
I would like to also thank you for your script. I used it also for cluster-FDR to see where the cut-off is (when using Ps I had a very smoothed curve with many values that was not very informative), but It did not help responding to the above question.
Kind Regards,
--
Maxime Résibois
PhD Student (KU Leuven)
Quantitative Psychology and Individual Differences
Tiensestraat 102 bus 3713
3000 LEUVEN
tel. +32 16 37 30 98
-----Original Message-----
From: Guillaume Flandin [mailto:[log in to unmask]]
Sent: Thursday 15 December 2016 12:37
To: Maxime Résibois <[log in to unmask]>
Cc: [log in to unmask]
Subject: Re: [SPM] Boundary of voxel-wise FDR correction
Dear Maxime,
I don't think there is a problem here and you are right that what you observe is related to the computation of the FDR q-value as min(QS(I:S) instead of QS(I) in spm_P_FDR.m.
Have a look at slide 41 for a visual description of how the Benjamini-Hochberg procedure works:
http://www.fil.ion.ucl.ac.uk/spm/course/slides14-may/05_ParamThresh.ppt
Depending on the profile of sorted uncorrected p-values, your could have a statistic with a larger uncorrected p-value being attributed a smaller FDR q-value.
To observe this with your data, run the following when looking at results (this is for peak-level topological FDR but you could display a similar plot for cluster- or voxel-level FDR):
QPp = xSPM.Pp;
QPp = sort(QPp(:));
S = numel(QPp);
figure, plot((1:S)/S,QPp,'-o',(1:S)/S,(1:S)/S*0.05);
xlim([0 1]);ylim([0 1])
set(gca,'YAxisLocation','right');
Best regards,
Guillaume.
On 14/12/16 09:41, Maxime Résibois wrote:
> Dear SPM experts,
>
>
>
> In case it might help finding where the issue might come from, I
> noticed that I encounter the same sort of results (i.e. exactly
> similar p-FDR
> values) also when using topological correction (and looking at the
> peak level, but also for two clusters in my first contrast when
> looking at the cluster level), and that this happens both in SPM 8 and SPM 12.
>
>
>
> Again, this seems to point to the spm_p_FDR.m way of selecting p-values.
>
>
>
> Any input would be greatly appreciated.
>
>
>
> Kind Regards,
>
> --
> *Maxime Résibois*
>
> PhD Student (KU Leuven)
> Quantitative Psychology and Individual Differences
> <https://ppw.kuleuven.be/okp/about_us/>
> Tiensestraat 102 bus 3713
> 3000 LEUVEN
> tel. +32 16 37 30 98
>
>
>
> *From:*Maxime Résibois
> *Sent:* Tuesday 13 December 2016 15:55
> *To:* SPM <[log in to unmask]>
> *Subject:* Boundary of voxel-wise FDR correction
>
>
>
> Dear SPM experts,
>
>
>
> I've encountered strange results concerning FDR correction.
>
> To start with, let me clarify that I refer to the voxel-wise (i.e.
> *non*-topological, with stats.topoFDR set to 0) FDR correction, and
> that I am only looking at the "peak" p-columns in SPM.
>
> So, when using an uncorrected threshold of p<.001 with k=10, resulting
> peak FDR values are very similar.
>
>
>
> In one contrast, of my 18 resulting peak, 17 FDR p-values are exactly
> 0.026131713.
>
> From what I have understood, FDR-correction is calculated from
> uncorrected p-values. However, the p(unc) leading to the same FDR
> results range from 0.00000398 and 0.00000744.
>
> Although they are /very /close, they are not exactly similar.
>
>
>
> In another contrast, of my 15 resulting peak, five are exactly
> 0.118333366538248, three exactly 0.138802207216991, two exactly
> 0.135174599683222 and two exactly 0.152021262602001. They are not in
> the same cluster.
>
> Again, p-values are very close but not exactly the same.
>
>
>
> I think it has to do with the way spm_P_FDR defines the FDR p-value
> (min(QS(I:S))
>
> I tried calculating the Qs by hand starting from the uncorrected
> p-values, but I could not.
>
>
>
> All in all, I do not understand why I get those strange results and
> could not find similar thread in mailing list.
>
> Would anyone have any idea of why those FDR /p/-values are that
> similar, and whether they are valid?
>
>
>
> Kind Regards,
>
>
>
> --
> *Maxime Résibois*
>
> PhD Student (KU Leuven)
> Quantitative Psychology and Individual Differences
> <https://ppw.kuleuven.be/okp/about_us/>
> Tiensestraat 102 bus 3713
> 3000 LEUVEN
> tel. +32 16 37 30 98
>
>
>
--
Guillaume Flandin, PhD
Wellcome Trust Centre for Neuroimaging
University College London
12 Queen Square
London WC1N 3BG
|