spatio_temporal_cluster_test with NaNs in data

Hey all,
new to MNE, and trying to use MNE’s spatio_temporal_cluster_test for some fNIRS data that has been separately processed using personal scripts (i.e., not using MNE etc).

As part of the pre-processing stage, artifacts are marked a NaNs, and I’m not sure what MNE/spatio_temporal_cluster_test is doing with the NaNs.

Artificially setting NaN to 0 across the data results in meaningful-looking results, but leaving NaNs in returns 0 clusters.

Is there any way of getting around having NaNs in the data?

For reference, I’m using MNE v 0.24.1 on Mac OS 12.3.1

Thanks in advance!

Hello @M055 and welcome to the forum!

I’m not quite sure where exactly you have NaNs in your data. Could you please give an example and maybe share a snippet of code that demonstrates what you’re doing?


Hello @richard
Sorry for the delay - I have NaNs because I have already-processed fNIRS [Oxy] data from an older dataset which I got as a 951(samples) X 24(channels) X 31(Subjects) from two conditions, and this data contains NaNs.

I’m basically trying to force previous non-MNE-data that contains NaNs into MNE’s stats.permutation_cluster_test function.

I did a little digging and I see that the nan_policy=‘omit’ is not yet implemented for the F-test in scipy, so it looks like my options might be to systematically remove NaNs from this data, or re-implement and pass an F-test that is ok with NaNs in the data.

I paste below a minimal sequence.


import numpy as np
from mne import stats

# Adjacency matrix from another file:
import NIRS_adj_mat
adj_matrix = NIRS_adj_mat.get_NIRS_adj_mat()

# Load the data from the obtained mat file
matdata ='./matlab_data/avg_keep.mat')

# Recover data for the 'R' and 'N' conditions
allOxyR = np.array(matdata['avg'][0])
allOxyN = np.array(matdata['avg'][1])

allOxyR = np.moveaxis(allOxyR,-1,0)
allOxyN = np.moveaxis(allOxyN,-1,0)

# Shape of arrays is 31(subjects) X 951(samples) X 24(channels)
allOxyR.shape, allOxyN.shape
>> ((31, 951, 24), (31, 951, 24))

# Make a copy where NaNs are replaced by 0

# Create an 'X' matrix with original data containing NaN and one with NaNs replaced by 0
X = [allOxyR, allOxyN]
X0 = [allOxyR0, allOxyN0]

# Run MNE clustering
cluster_stats = stats.permutation_cluster_test(X,adjacency=adj_matrix)
>> Using a threshold of 4.001191
>> stat_fun(H1): min=nan max=nan
>> Running initial clustering
>> Found 0 clusters

cluster_stats = stats.permutation_cluster_test(X0,adjacency=adj_matrix)
>> Using a threshold of 4.001191
>> stat_fun(H1): min=nan max=nan
>> Running initial clustering
>> Found 42 clusters
>> Permuting 1023 times...