Ok... So, the thing I was getting wrong was the effect of the df, which I wasn't considering.
With mice N is generally very small, and apparently this massively impacts the threshold. The mice brain mask is more spherical than the human one, and it doesn't have holes (for ventricles) in it.
I've tried using use spm_uc_RF assuming df = [1 100] and apparently the result is 3.6 for 6.8 resels (mice narrow mask), 4.2 for 67 resels (mice wider mask, smaller FWHM) and t=4.8 for 636 resels. I guess that the impact of the number of multiple comparisons isn't that big on the exact threshold. I guess it boils down to the fact that when you're in the tail of the Gaussian the order of magnitude of p changes a lot with small changes in t!
Best regards,
Luca
-----Messaggio originale-----
Da: Flandin, Guillaume [mailto:[log in to unmask]]
Inviato: venerdì 29 novembre 2019 14:05
A: PRESOTTO LUCA <[log in to unmask]>; [log in to unmask]
Oggetto: Re: R: [SPM] Wrong multiple comparisons corrections in mice
Dear Luca,
I indeed assumed you already knew how RFT works but was confused by your comments on brain sizes. You can see what happens by directly calling the function that returns the corrected critical height threshold:
u = spm_uc_RF(a,df,STAT,R,n)
where a is 0.05, df is [1 df], STAT is 'T', R is SPM.xVol.R and n is 1.
RFT will actually depend on the geometry of the search volume (the 'volume' in each dimension), i.e. SPM.xVol.R is a vector and not a scalar, and the penalty gets stronger the more you get away from a spherical shape. Could that be the issue for you here?
Best regards,
Guillaume.
On 29/11/2019 12:33, PRESOTTO LUCA wrote:
> Dear Guillaume,
> I was aware of this concept, even if I always mess up the details. My doubt was indeed relative to what SPM is getting wrong about the RESEL. I was giving the spatial details to imply that these brains are extremely small compared to the resolution, therefore I expect there to be something like ~10 multiple comparisons in total!
> So, on a human brain smoothed 8 mm I get 635 resels in the summary. In
> my mice brain I get between 6 and 60 resels (depending on the mask size and on the estimated FWHM, that varies wildly for reasons between cases....) Given that there are so few resels shouldn't the FWE correction be almost negligible, instead of resulting in t>9?
>
> Best Regards,
> Luca
>
> -----Messaggio originale-----
> Da: Flandin, Guillaume [mailto:[log in to unmask]]
> Inviato: venerdì 29 novembre 2019 12:16
> A: PRESOTTO LUCA <[log in to unmask]>; [log in to unmask]
> Oggetto: Re: [SPM] Wrong multiple comparisons corrections in mice
>
> Dear Luca,
>
> Roughly speaking, correction for multiple testing with the random field theory will take into account the number of tests (i.e. voxels) and their spatial correlation (i.e. smoothness), summarised with the RESEL count. The actual size of the search volume (in mm) does not really matter and you could have a stronger penalty for a small brain compared to a bigger brain if the RESEL count is larger in the former compared to the later. See e.g. slide 26 here:
>
> https://www.fil.ion.ucl.ac.uk/spm/course/slides18-oct/05_Thresholding.
> pptx You can see the values used by SPM in SPM.xVol, and in particular
> fields FWHM (smoothness), R (RESEL count) and S (# voxels).
>
> Best regards,
> Guillaume.
>
>
> On 28/11/2019 18:08, PRESOTTO LUCA wrote:
>> Dear experts,
>>
>> I'm implementing my PET mice brain analysis in SPM. It works kind of
>> well but the FWE correction gives unreasonable results. Considering
>> how small the brains are I expect that it shouldn't have much impact.
>> Instead I find corrections that look for t>5/t>9. I can't exactly
>> point out what's going wrong. I can see that sometimes it
>> overestimates the smoothness (FWHM 2.6 - 4.5 mm... I've smoothed the
>> images to 1.8 mm). My voxel size is 0.2 mm (isotropic).
>>
>> Anyway... Considering how small a mouse brain is, compared to even
>> just
>> 2.0 mm resolution I would expect almost no impact from the FWE
>> correction. Instead if goes up to t>9!! It's 4.5 with a human brain!
>>
>> But I can't clearly pinpoint a specific error anywhere... Any
>> suggestion on how to debug what's going on?
>>
>> I'm using an explicit mask, which happens to be 33k voxels big, if
>> this matters. The human mask I use is 236k big!!
>>
>>
>>
>> Thank you in advance!
>>
>> Luca Presotto
>>
>>
>>
>> <https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fso
>> s
>> tienici.hsr.it%2Fsostienici%2Fquesto-natale-sostieni-la-ricerca-insie
>> m
>> e-a-noi.html&data=02%7C01%7C%7C029c3f7ba8884bfe3ef908d7742dfe4f%7C1fa
>> f
>> 88fea9984c5b93c9210a11d9a5c2%7C0%7C1%7C637105613195360469&sdata=xsdnw
>> I V8VIJFm6zriHk6K7vAuWaavFhewNzwso%2BZUAc%3D&reserved=0>
>>
>>
>>
>>
>> /Rispetta l'ambiente: non stampare questa mail se non è necessario.
>> Respect the environment: print this email only if necessary./
>>
>
> --
> Guillaume Flandin, PhD
> Wellcome Centre for Human Neuroimaging UCL Queen Square Institute of
> Neurology London WC1N 3BG
> [https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fpor
> talepaziente.hsr.it%2FNatale2019_BannerFirmaMail.jpg&data=02%7C01%
> 7C%7C08b084f19c8e4d6803d708d774c84efe%7C1faf88fea9984c5b93c9210a11d9a5
> c2%7C0%7C0%7C637106275975448609&sdata=MQy1HexWZvH3kUmB0q51lYZbj%2B
> 6SSfXyAMNAyLqINiE%3D&reserved=0]<https://eur01.safelinks.protectio
> n.outlook.com/?url=https%3A%2F%2Fsostienici.hsr.it%2Fsostienici%2Fques
> to-natale-sostieni-la-ricerca-insieme-a-noi.html&data=02%7C01%7C%7
> C08b084f19c8e4d6803d708d774c84efe%7C1faf88fea9984c5b93c9210a11d9a5c2%7
> C0%7C0%7C637106275975448609&sdata=mCndJF0a%2FiSTHkNQ1Vw%2BCG4ZbjZF
> yWqiOF93gfioaV0%3D&reserved=0>
>
--
Guillaume Flandin, PhD
Wellcome Centre for Human Neuroimaging
UCL Queen Square Institute of Neurology
London WC1N 3BG
[https://portalepaziente.hsr.it/Natale2019_BannerFirmaMail.jpg]<https://sostienici.hsr.it/sostienici/questo-natale-sostieni-la-ricerca-insieme-a-noi.html>
Rispetta l’ambiente: non stampare questa mail se non è necessario.
Respect the environment: print this email only if necessary.
|