Threshold_stats_img fpr vs fdr correction


Can someone please explain the difference between frd and fpr corrections in nilearn.glm.threshold_stats_img ?

Thanks a lot in advance!


  • ‘fpr’ controls the rate of false positive rate. For instance, a fpr control=0.001, means that, among all null voxels, a fraction of 0.1% will falsely declared active. We don’t know this number, but it is typically close to the total number of voxels. So you can expect the number of false positives to be .001 * #voxels
  • ‘fdr’ controls the [expected value of] the rate of false positives among detections. This means that for a FPR of 10% you can roughly consider that 90% of the detections are true positive, 10% are false positives.
    A typical value for fpr (without correction for multiple comparisons) is .001
    A typical value for fdr control is 5 or 10%.
    I would advise to use fdr control.
1 Like