Incredibly high p-value

#1

Hi MRtrix experts,

I am running fixelcfestats command with sex, age, motion, whole brain volume and IQ. The result gives me a incredibly high p-value that cannot be true (1.17549435e-38). Once I run it without whole brain volume, it gives a reasonable p-value, so it seems the problem lies somehow with the brain volume values. Both FWE-corrected and uncorrected p-values are high. Do you see anything which should be adjusted?

GLM design: [1, sex, age, motion, brain volume, IQ (all centered score except for 1s and sex)]
1 1 9.684 0.086 -13755.116 -2.34751773
1 0 9.684 -0.413 -3381.216 0.65248227
1 1 4.684 -0.218 -111982.916 -1.34751773
1 0 10.684 0.802 -252771.016 -0.34751773
1 1 -5.316 -0.079 104972.084 3.65248227
1 1 -7.316 0.062 46757.484 -5.34751773
1 0 8.684 -0.219 -157965.516 -2.34751773
…

Code:
fixelcfestats -force -negative -nperms 5000 $DIR/fc/ $DIR/scripts_fixelcfestats/input_281.txt $DIR/scripts_fixelcfestats/design_ones_sex_age_motion_WBV_wisc.txt $DIR/scripts_fixelcfestats/contrast_matrix_000001.txt $DIR/tracks_2_million_sift.tck $DIR/fixelcfe_output/fixelcfestats_ones_sex_age_motion_WBV_1wais_fc

Result:

image
image.jpg902×789 95.3 KB

Thank you for your help in advance!
Yuki

#2

Hi @yuki,

How are you specifying the statistical model (i.e. the GLM contrast)?

Could elaborate on your effects-of-interest, and those which you want to covary for?

Best,

Alistair

#3

The result gives me a incredibly high p-value that cannot be true (1.17549435e-38).

Describing a couple of confounds here just in case:

  • 1e-38 is in fact a very small number: it’s a decimal point followed by 37 zeroes. Moreover, the image likely doesn’t actually contain this value; what’s more likely happening is that the image consists entirely of zeroes, but mrview is setting the upper limit of the colour bar to the smallest number it can represent that is greater than zero (the lower limit of the colour bar), in order to avoid dividing by zero during various calculations. This could be confirmed by running mrstats on the fixel data file, and checking the minimum and maximum values.

  • The “p-value” images provided by statistical inference commands in MRtrix3 in fact provide values as (1.0-p). So when the value stored in image fwe_pvalue.mif is ~ 0.0, this in fact represents p ~ 1.0, i.e. no effect.

So your statement that “the p-values are high” is strictly true; just making sure it wasn’t due to two errors cancelling one another out :stuck_out_tongue:

If this is the case, the data you need to be looking at more closely are the contents of image cfe.mif and file perm_dist.txt, to see if the enhanced t-statistic values in the default permutation are in fact much smaller than those observed in the null distribution.

Based on recent experience, I would also suggest importing your design matrix into Matlab or Octave and calculating its “condition number”. If this is >> 1.0 it can lead to all sorts of issues…