Nilearn bonferroni correction

1

Hello,

I am trying out Nilearn for fMRI analysis.
In my group analysis, I get a z_map

z_map = second_level_model.compute_contrast(output_type=‘z_score’)

Now I want to get a FWE corrected threshold. I do this:

_, threshold = threshold_stats_img(
z_map, alpha=0.05,
height_control=‘bonferroni’,
cluster_threshold=0,
two_sided=False)

I wonder, is Nilearn doing this bonferroni correction on the number of voxels or on the number of resels?

2

On voxels.
Best,
Bertrand

3

Thanks, Bertrand!
Yeah, I figured that too, looking at the source code - Bonferroni threshold is calculated as an ‘inverse survival function’ (inf) of alpha/n_voxels. And that returns a z-score.
And apparently is done like that independent on what is the input image stat_img in the threshold_stats_img. So, even if my stat_img is t-value image or F-value image, the returned threshold is still a z-score for alpha/n_voxels. And applying that z-score threshold to F-value image is just wrong.
That’s quite dissapointing. And also in this way thresholding fMRI images is rather useless (correcting for the number of voxels rather than resels). :frowning:

Are there any other Python based fMRI statistical analysis tools available?

4

Interesting point.
To me it would not make sense to create Bonferroni control for all possible statistical decision function. It makes more sense to do it for z-values only. What you are saying is that F-to-z and t-2-z converters are lacking. We would indeed add them.
(Currently, Nilearn GLMs directly return z-scale variables ; the point is that bare t/F statistical maps are not interpretable).
Best,
Bertrand

5

It is just not clear that nilearn.glm.threshold_stats_img(stat_img) implies that the stat_img should always be z-scaled image. It’s missleading. I assumed it can be any stats image.

I have now discovered that Nilearn also has a non_parametric_inference module. That’s at least helpful to overcome the lack of RFT implementation!

Thanks!

6

It is in the doctring. nilearn.glm.threshold_stats_img - Nilearn
Do you think it should be more explicit ?
Glad that you could indeed use non-parametric inference. This is certainly a good solution —albeit more expensive.
Best,
Bertrand