 |
 |
Hi Anderson, Thank you for this detailed and clear explanation. Best regards, Matthieu 2016-06-06 8:28 GMT+02:00 Anderson M. Winkler <[log in to unmask]>: > Hi Matthieu, > > Probably the simplest way to obtain the residuals is with the command > fsl_glm (option --out_res). > > About whether PET data is symmetric: even if symmetry cannot be rejected > in a particular dataset (i.e., KS test not significant), I see no reason in > your analysis to introduce this assumption, even if the point were to > obtain a larger number of possible shufflings. There are 3003 possible > permutations with the smallest groups in your data, which should be plenty, > and there are yet far more with the other groups. > > All the best, > > Anderson > > > On 5 June 2016 at 12:32, Matthieu Vanhoutte <[log in to unmask]> > wrote: > >> Hi Anderson, >> >> Please see below: >> >> Le 5 juin 2016 11:42, "Anderson M. Winkler" <[log in to unmask]> a >> écrit : >> > >> > Hi Matthieu, >> > >> > Please, see below: >> > >> > >> > On 4 June 2016 at 14:01, Matthieu Vanhoutte < >> [log in to unmask]> wrote: >> >> >> >> Hi Anderson, >> >> >> >> Please see below: >> >> >> >> On 06/04/2016 12:07 PM, Anderson M. Winkler wrote: >> >>> >> >>> Hi Matthieu, >> >>> >> >>> The skewness of the residuals depends on the data and on the >> variables in the model, although the variables often don't remove all >> skewness if that is present in the original data. It's something that >> generally goes by assumption, i.e., it doesn't have to be tested in all and >> every study, as the modalities we use are generally members of a not too >> large set. FMRI and cortical thickness, for instance, can probably be >> safely assumed as symmetric, whereas VBM, FA, and surface area cannot. >> >> >> >> >> >> Have you any papers that would confirm that FMRI and cortical >> thickness can probably be safely assumed as symmetric, whereas VBM, FA, and >> surface area cannot ? Any other papers concerning supplementary data types ? >> > >> > >> > It's simpler to reject a hypothesis than confirm it. We know for sure >> that VBM is skewed, and so is FA. The distribution for area depends on the >> resolution: the finer, the more log-normally distributed (so, more skewed). >> For other data, there is always a possibility that the distribution is >> asymmetric (hence obviously not normal), hence permutation tests are so >> important and should be favoured. >> > >> > >> >>> >> >>> >> >>> At any rate, to test symmetry, a simple option is to take the >> residuals of the unpermuted data, make a copy with all signs flipped, and >> do a two-sample Kolmogorov-Smirnov test (a test that compares two >> distributions). >> >> >> >> >> >> Is there a simple way of doing this 3 steps above with palm or any >> other software ? >> > >> > >> > Yes, the KS test is available in Matlab with the Statistics toolbox >> (command kstest), and in Octave (command kolmogorov_smirnov_test). It's >> also available in R and in Python (w/ scipy). Interestingly, it's possible >> to put the test in a for-loop and do a simple permutation test here too, >> that is, a permutation test to see if the distribution is symmetric. >> >> How could I get the residuals of the unpermuted data ? Sorry but I don't >> well understand how to see if the distribution is symmetric... >> >> > >> > It seems any of this won't be needed, though, please see below: >> > >> >>> >> >>> >> >>> Perhaps more important is that more shufflings (permutations and/or >> sign-flippings) do not increase power, and in fact, using sign-flippings >> when the distribution is skewed causes the test to be conservative (See >> Figures 1 and 2, and Table 2 of the Supplementary Material of this paper). >> >>> >> >>> How many subjects are there in each sample? How many permutations are >> possible? What type of imaging data are you using? >> >> >> >> >> >> 1) I have 4 groups with respectively 29, 5, 10 and 10 subjects. >> >> 2) How many permutations are possibles are defined in the Neuroimage >> 2014 paper ? But practically, using -n 5000 is a good way, isn't it ? >> > >> > >> > This is a quite large sample. The overall (all groups) number of >> possible unique permutations is (29+5+10+10)!/29!/5!/10!/10!, which is a >> big number. No need to add sign-flippings so as to increase the number of >> shufflings (and thus incurring the risk of violating symmetry assumptions). >> > >> > If you run a contrast comparing only the smallest groups (5 and 10), >> the number of possible permutations here is smaller, 15!/10!/5! = 3003, but >> still ok, and the result is exact. No need for sign-flippings here either. >> >> Ok thanks ! >> >> > >> > All the best, >> > >> > Anderson >> > >> > >> >> >> >> 3) I am using FDG-PET data >> >> Any prior knowledge to know about this kind of data ? >> >> Best regards, >> Matthieu >> >> >> >> >> Hope you could give me any advice ! >> >> >> >> Best regards, >> >> Matthieu >> >> >> >>> >> >>> All the best, >> >>> >> >>> Anderson >> >>> >> >>> >> >>> On 3 June 2016 at 21:53, Matthieu Vanhoutte < >> [log in to unmask]> wrote: >> >>>> >> >>>> Dear FSL experts, >> >>>> >> >>>> I'd like to compare 4 groups of patients with palm. I have looked at >> the paper explaining permutation inference from Neuroimage (2014), but >> would like a personal advice about use of -ee and -ise. >> >>>> >> >>>> Classically I would use the -ee option for my study, but why not >> -ise at the same time to increase statistical power un case of small groups >> ? >> >>>> >> >>>> How to determine that in a model the errors are symmetric ? >> >>>> >> >>>> Best regards, >> >>>> >> >>>> Matthieu >> >>> >> >>> >> >> >> > >> > >