The problem points are the super-high intensity points in the RPV image.  Those super-rough regions are artificially shrinking your FWHM estimate which is sending your RESEL count through the roof (can you report what your FWHM & Resel counts are?)

The solution is to mask out (better yet, understand what's going on with) those super-rough voxels.  Worst case, take the mask image created by SPM, and then mask out the voxels with super-high RPV values (with some ImCalc arithmatic) and create a new mask that you can uses as an 'explicit mask' in a new analysis.

Hope this helps!


On Wed, Oct 8, 2008 at 8:43 PM, Carlos Faraco <[log in to unmask]> wrote:
On Mon, 6 Oct 2008 21:51:26 +0100, Thomas Nichols <[log in to unmask]>

>Are the problems related to smoothness estimation?

Do you mean at the individual level or 2nd level? How can I see values for this?

>I.e. is the smoothness FWHM or Resels NaN as well?  Or, are their strange
values for the
>"Expected voxels per cluster" or "Expected number of clusters"?  Is the
number of the
>voxels in the analysis what you expect?  (Mask not way too large or too

On the within group data I receive the following output:

For Stroop task, using default values on results:
1. Using no correction: some clusters, Expected voxels per cluster (EVC) =
.003, Expected # of clusters (E#C) = 64075.98, Expected FDR = .66
2. Using FDR: No suprathreshold clusters, EVC = NaN, E#C = NaN, EFDR = .66
3. Using FWE: 2 clusters, 1 large one with 199336 voxels, EVC = 221, E#C =
.05, FDR = .90

For OSPAN task:
1. No correction: Showed scattered, random looking activation and gave the
following warning many times
Warning: Returning NaN for out of range arguments
> In spm_Pcdf at 88
 In spm_P_RF at 106
 In spm_P at 48
 In spm_list at 487

I pressed Ctrl+C to stop the repition of these warnings and the volume table
had a 25 page list of clusters all with NaNs for Pcorrected.

2.Using FDR: same as in Stroop, except EFDR = .10

3. Using FWE: no clusters, EVC=0.000, E#C = 393619779.31, EFDR=.10

The # of voxels for the analyses were around  205,000 voxels.

>If there are NaN's in the smoothness estimate, look at the FWHM image
>(compute from the RPV image with ImCalc as FWHM = RPV.^(1/3)).  Viewing this
>image with CheckReg with other images (e.g. the mask, or T image, ResMS (or
>its square root), etc).  If the is *really* irregularly shaped (i.e. no
>contiguous voxels) it is possible to have almost all NaN's in the RPV image.

The RPV images for the group data have some NaN values outside of the brain,
which as you suggested may be normal. However, the images are of a very low
intensity, from about  .002 - 3.0, expect for a few high intensity points.
The FWHM from this doesn't show much of course.

I hope this helps to describe the problem. Also, some of the activation on
the single subject data looks really bad, just big huge blobs.


>On Mon, Oct 6, 2008 at 9:08 PM, Carlos Faraco <[log in to unmask]> wrote:
>> Dear SPMers,
>> Today I read through many of the posts regarding NaNs, however I am still
>> confused as to what would be the best remedy for my situation.
>> The problem is I keep on receiving NaNs for the p-vals on the second level
>> analysis. I tried changing the NaNs to 0s in the con images (which seem to
>> appear only or mostly outside the brain; and becasue I read not to do this
>> for the beta images) using ImCalc but that still doesn't work. I therefore
>> checked some files from some other studies we had previously done, and it
>> appears that setting voxels outside the brain to NaN is normal SPM
>> behavior.
>> Is the fix for this then something that has to be modified within the SPM
>> code?
>> Thanks,
>> Carlos Faraco
>Thomas Nichols, PhD
>Director, Modelling & Genetics
>GlaxoSmithKline Clinical Imaging Centre
>Senior Research Fellow
>Oxford University FMRIB Centre

Thomas Nichols, PhD
Director, Modelling & Genetics
GlaxoSmithKline Clinical Imaging Centre

Senior Research Fellow
Oxford University FMRIB Centre