spectral_connectivity.statistics.Benjamini_Hochberg_procedure#

Benjamini_Hochberg_procedure(p_values: ndarray[tuple[int, ...], dtype[floating]], alpha: float = 0.05) → ndarray[tuple[int, ...], dtype[bool]][source]#

Control false discovery rate using Benjamini-Hochberg procedure.

Corrects for multiple comparisons and returns significant p-values by controlling the false discovery rate at level alpha using the Benjamini-Hochberg procedure.

Parameters:

  • p_values (NDArray[floating], shape (...,)) – P-values from statistical tests to be corrected.

  • alpha (float, default=0.05) – Expected proportion of false positive tests (false discovery rate).

Returns:

is_significant – Boolean array same shape as p_values indicating whether the null hypothesis has been rejected (True) or failed to reject (False).

Return type:

NDArray[bool], shape (…,)

Examples

>>> import numpy as np
>>> p_vals = np.array([0.001, 0.02, 0.04, 0.3, 0.8])
>>> significant = Benjamini_Hochberg_procedure(p_vals, alpha=0.05)
>>> significant
array([ True,  True, False, False, False])

Notes

Non-finite p-values (NaN/inf) mark undefined tests — for example a coherence pair involving a dead/zero-power channel — and are excluded from the family: they neither count toward the number of tests nor tighten the threshold for the valid ones, and are returned as False. If every p-value is non-finite (the whole family is undefined, e.g. every tested pair involves a dead channel) a UserWarning is emitted, because the all-False result would otherwise be indistinguishable from a valid family with no true effects. All input dimensions are pooled into a single family (the array is raveled); the result has the input shape. Delegates to scipy.stats.false_discovery_control() (SciPy >= 1.11; a clear RuntimeError is raised on older SciPy), which rejects finite p-values outside [0, 1].