def compute_pval_rsa(seed):
stim, voxels = load_data(n_samples, n_features, model=model, seed=seed,
heteroscedastic=heteroscedastic)
# compute similarity
stim_ = stim
if stim.shape[1] == 1:
stim_ = np.hstack((stim, - stim))
stim_similarity = square_pdist(stim_) # np.corrcoef(stim_)
voxels_similarity = square_pdist(voxels) # np.corrcoef(voxels)
# indices to extract lower triangular part of a matrix
lw_idx = np.triu_indices(n_samples, k=1)
stim_vsim = stim_similarity[lw_idx]
voxels_vsim = voxels_similarity[lw_idx]
# compute the statistic
# T = np.corrcoef(stim_vsim, voxels_vsim)[0, 1]
T = spearmanr(voxels_vsim, stim_vsim)[0]
T_perm = []
for i in range(n_draws):
# permute the labels
perm = np.random.permutation(n_samples)
# voxels_vsim_perm = np.corrcoef(voxels[perm])[lw_idx]
voxels_vsim_perm = square_pdist(voxels[perm])[lw_idx]
# compute the test statistic
# T_perm.append(np.corrcoef(voxels_vsim_perm, stim_vsim)[0, 1])
T_perm.append(spearmanr(voxels_vsim_perm, stim_vsim)[0])
pval = 1 - percentileofscore(np.array(T_perm), T) / 100.
return pval
#
# True: Both models perform equally well
# False: Linear models outperform RSA
pvals_rsa = []
a_space = np.linspace(0, .3, 11)
for a_i in a_space:
T_ai = []
# average across repetitions
for i in range(n_repeats):
stim, voxels = load_data(
n_samples, n_voxels, activation=a_i, model=model, seed=i,
random_effects=random_effects)
# compute similarity
stim_similarity = (stim - stim.T) ** 2
voxels_similarity = - square_pdist(voxels) # np.corrcoef(voxels)
# extract lower triangular part of symmetric similarity
lw_idx = np.triu_indices(n_samples, k=1)
stim_vsim = stim_similarity[lw_idx]
voxels_vsim = voxels_similarity[lw_idx]
# compute the statistic
T_ai.append(spearmanr(stim_vsim, voxels_vsim)[0])
T_perm = []
for _ in range(n_perm):
# permute the labels
perm = np.random.permutation(n_samples)
voxels_vsim_perm = - square_pdist(voxels[perm])[lw_idx]
