Dear Jason,
After doing some more extensive tests on fdr today I have
come across one instance where it does something
different from the matlab code, although this doesn't really
seem to be your case. If you do not use a mask then it
creates an implicit mask by thresholding at p<0.9999
which can exclude some voxels and cause differences
between the matlab code and the FSL code.
However, if you do specify a mask then I get exactly
the same values from matlab and fsl. Can you double
check that you really do see a difference and if it still
exists then send us the images (to our upload site) so
that we can try to see what is going on.
All the best,
Mark
On 17 Sep 2009, at 09:07, Mark Jenkinson wrote:
> Hi,
>
> Did you get nothing for both pID and pN in FDR.m ?
> Can you run your fdr command with the -v option and send us
> the output? I'd also like to know the size of the mask (fslstats
> with -V
> will tell me what I want to know).
>
> As your threshold is very low, I'm assuming that not much passes
> this, but I cannot quite see why the results are different.
>
> All the best,
> Mark
>
>
>
> On 17 Sep 2009, at 01:58, Jason Stein wrote:
>
>> Hello FSLers,
>>
>> I have a question regarding the FSL fdr command.
>>
>> First, I created a map of P-values from a program outside of FSL
>> which does a multiple linear regression. Then, I created a
>> histogram of the P-values within a mask of interest. This
>> histogram is not attached due to space limits (my previous email
>> was rejected). A description is that there are many voxels with
>> high P-values and fewer voxels with low P-values (a left half
>> frowning histogram). Then, I ran two FDR corrections. (a) Taking
>> only the P-values within the mask using Matlab I used the FDR.m
>> script from Tom Nichols' website, and (b) The fdr command in FSL.
>>
>> Method (a) gave me no P-value that passes the FDR at a q-value of
>> 0.05 using the standard Benjamini Hochberg 1995 FDR method.
>> However, method (b) gave me a P-value threshold - it did pass.
>> Considering that the histogram looks the opposite of what I would
>> expect for passing FDR and that methods (a) and (b) give different
>> results, I'm confused as to what's going on. Can someone help?
>>
>> Also, the histogram of P-values we were studying were created by
>> studying FA differences using standard parametric regression. Is
>> the reason the histogram looks funny (in that a null histogram
>> should be flat, not having fewer voxels with less significance)
>> because we are not satisfying the distributional assumptions of a
>> parametric regression? In other words, should we be using non-
>> parametric regression?
>>
>> Thanks in advance for any help,
>> Jason Stein
>>
>>
>>
>> fdr -i /Users/M/Desktop/Stats/Output_stats/EFFECT_V/
>> VolumeRegressionAnalysis.OutputP-values_0-2.img -m /Users/M/Desktop/
>> Stats/Mori_mask_WM_stats_220.img -q 0.05
>> Probability Threshold is:
>> 0.000392611
>>
>>
>>
>
|