Hi everyone!
- MNE version: MNE 1.9.0
- macOS 15.6
I am doing fully exploratory analysis and have been trying to run a permutation cluster test with a 2 (group: patients x HC) x 3 (task: task1 x task2 x task3) ANOVA, instead of a simple bivariate comparison.
I have run the one-way F-test for each group separately. There is an effect for two tasks but not for one. This requires a test of interaction. I initially tried to do this based on differences per group (e.g. task1-task2), but this does not give me meaningful results and does not find any significant clusters at all, no matter what the threshold.
So I figured, I better run a real ANOVA instead. I used the following code, but I could not progress due to errors that I could not solve.
# -------------------------------
# Define groups (HC = 1, SSD = 0)
# -------------------------------
groups = np.array(
[1]*hc_group_data_for_spatiotemporal_test["ec"].shape[0] +
[0]*ssd_group_data_for_spatiotemporal_test["ec"].shape[0]
)
n_subjects = len(groups)
print("Groups array:", groups)
# -------------------------------
# Prepare X_list for all 3 conditions (ec, eo, hct)
# -------------------------------
X_list = []
for task in ["ec", "eo", "hct"]:
X_hc = hc_group_data_for_spatiotemporal_test[task] # shape: n_HC × n_channels × n_times
X_ssd = ssd_group_data_for_spatiotemporal_test[task] # shape: n_SSD × n_channels × n_times
X_all = np.concatenate([X_hc, X_ssd], axis=0) # n_subjects_total × n_channels × n_times
# Transpose for MNE: (n_subjects, n_times, n_channels)
X_all = np.transpose(X_all, (0, 2, 1))
X_list.append(X_all)
print("Prepared X_list shapes (subjects, times, channels) for each condition:")
for cond_idx, arr in enumerate(X_list):
print(f"{['ec','eo','hct'][cond_idx]}: {arr.shape}")
# -------------------------------
# Build combined adjacency (spatial + temporal)
# -------------------------------
# Use one evoked from HC as template
template_evoked = list(hc_data['ec'].values())[0][0] # first subject, first evoked
channels_of_interest = list(range(0, 31)) # adjust as needed
# Spatial adjacency
adjacency, ch_names = find_ch_adjacency(template_evoked.info, ch_type='eeg')
ch_idx = channels_of_interest
adjacency = adjacency[ch_idx, :][:, ch_idx]
# Temporal adjacency: 1-neighbor adjacency along time
n_times = X_list[0].shape[1]
temporal_adj = np.eye(n_times, k=1) + np.eye(n_times, k=-1)
# Combine adjacency
combined_adj = combine_adjacency(adjacency, temporal_adj)
print("Combined adjacency shape:", combined_adj.shape)def mixed_anova_stat_fun_multi_channel(X, groups):
# X is a tuple of arrays (one per condition)
n_conditions = len(X)
n_subjects = X[0].shape[0]
# Infer number of times and channels from total features
n_features = X[0].shape[1]
n_times = 75
n_channels = 31
# Reshape each condition back to (subjects, times, channels)
X_reshaped = [x.reshape(n_subjects, n_times, n_channels) for x in X]
F_vals_condition = np.zeros((n_times, n_channels))
F_vals_interaction = np.zeros((n_times, n_channels))
for t in range(n_times):
for ch in range(n_channels):
data = np.column_stack([cond[:, t, ch] for cond in X_reshaped])
# Create a dataframe for mixed ANOVA
subjects = np.arange(n_subjects)
df = pd.DataFrame({
'y': data.flatten(),
'condition': np.repeat(['ec', 'eo', 'hct'], n_subjects),
'subject': np.tile(subjects, n_conditions),
'group': np.tile(groups, n_conditions)
})
aov = pg.mixed_anova(dv='y', within='condition', between='group',
subject='subject', data=df)
F_vals_condition[t, ch] = aov.loc[aov['Source'] == 'condition', 'F'].values[0]
F_vals_interaction[t, ch] = aov.loc[aov['Source'] == 'Interaction', 'F'].values[0]
# Stack along a new axis: 0 = condition effect, 1 = interaction
F_vals = np.stack([F_vals_condition, F_vals_interaction], axis=0) # shape (2, n_times, n_channels)
return F_vals
# run cluster perm test
F_obs_condition, F_obs_interaction, clusters, cluster_pv, H0 = spatio_temporal_cluster_test(
X_list,
adjacency=combined_adj,
n_permutations=1000,
threshold=None,
tail=0,
stat_fun=lambda *X: mixed_anova_stat_fun_multi_channel(list(X), groups),
verbose=True
)
# Error I get:
ValueError Traceback (most recent call last)
Cell In[532], line 4
1 # -------------------------------
2 # Run spatio-temporal cluster test
3 # -------------------------------
----> 4 F_obs_condition, F_obs_interaction, clusters, cluster_pv, H0 = spatio_temporal_cluster_test(
5 X_list,
6 adjacency=combined_adj,
7 n_permutations=1000,
8 threshold=None,
9 tail=0,
10 stat_fun=lambda *X: mixed_anova_stat_fun_multi_channel(list(X), groups),
11 verbose=True
12 )
15 print("Cluster test finished!")
16 print("F_obs_condition shape:", F_obs_condition.shape)
File <decorator-gen-306>:10, in spatio_temporal_cluster_test(X, threshold, n_permutations, tail, stat_fun, adjacency, n_jobs, seed, max_step, spatial_exclude, step_down_p, t_power, out_type, check_disjoint, buffer_size, verbose)
File ~/miniconda3/envs/mne/lib/python3.12/site-packages/mne/stats/cluster_level.py:1554, in spatio_temporal_cluster_test(X, threshold, n_permutations, tail, stat_fun, adjacency, n_jobs, seed, max_step, spatial_exclude, step_down_p, t_power, out_type, check_disjoint, buffer_size, verbose)
1552 else:
1553 exclude = None
-> 1554 return permutation_cluster_test(
1555 X,
...
---> 13 X_reshaped = [x.reshape(n_subjects, n_times, n_channels) for x in X]
15 F_vals_condition = np.zeros((n_times, n_channels))
16 F_vals_interaction = np.zeros((n_times, n_channels))
ValueError: cannot reshape array of size 103000 into shape (103,75,31)
Maybe you have a solution or another idea as to how I should approach this analysis?
I would really appreciate it!
Cheers,
Deniz
).