def nistats_confound_glm(nifti_file, confounds_file, which_confounds):
import pandas as pd
import nibabel as nb
import numpy as np
from nistats.regression import OLSModel
from scipy.stats import zscore
from nilearn.image import math_img
from nilearn.plotting import plot_stat_map
import matplotlib.pyplot as plt
infile = nb.load(nifti_file)
mean_img = math_img('np.mean(infile, axis=-1)', infile=infile)
in_data = infile.get_data().astype(np.float32)
confounds_table = pd.read_table(confounds_file)[which_confounds]
# we assume the confounds nor the nifti_file need temporal filtering
cfz = confounds_table.apply(zscore)
cfz['intercept'] = np.ones(infile.shape[-1])
om = OLSModel(np.array(cfz))
om_rr = om.fit(in_data.reshape((-1,infile.shape[-1])).T)
resid_img = nb.Nifti1Image(om_rr.resid.T.reshape(infile.shape).astype(np.float32), affine=infile.affine, header=infile.header)
cleaned_img = math_img('(resid_img + mean_img[...,np.newaxis]).astype(np.float32)', resid_img=resid_img, mean_img=mean_img)
output_nifti = nifti_file.replace('.nii.gz', '_nuis.nii.gz')
cleaned_img.to_filename(output_nifti)
output_pdf = confounds_file.replace('.tsv', '_sd-diff.pdf')
f = plt.figure(figsize=(24,6))
plot_stat_map(math_img('(infile.std(axis=-1)-cleaned_img.std(axis=-1))/infile.std(axis=-1)', infile=infile, cleaned_img=cleaned_img),
bg_img=mean_img, figure=f, cut_coords=(0,0,0), threshold=0.125, vmax=1, cmap='viridis', output_file=output_pdf)
return output_pdf, output_nifti
def test_residuals():
Xintercept = X.copy()
# If design matrix contains an intercept, the
# mean of the residuals should be 0 (short of
# some numerical rounding errors)
Xintercept[:, 0] = 1
model = OLSModel(design=Xintercept)
results = model.fit(Y)
assert_almost_equal(results.residuals.mean(), 0)
assert len(results.whitened_residuals) == 40
def test_predicted_r_square():
Xshort = X.copy()[:10, :]
Yshort = Y.copy()[:10]
# Signal of 10 elements should be completely
# predicted by 10 predictors (short of some numerical
# rounding errors)
model = OLSModel(design=Xshort)
results = model.fit(Yshort)
assert_almost_equal(results.residuals.sum(), 0)
assert_array_almost_equal(results.predicted, Yshort)
assert_almost_equal(results.r_square, 1.0)
def test_resid_rename_warnings_ols():
model = OLSModel(design=X)
results_ols = model.fit(Y)
with pytest.warns(FutureWarning, match="'df_resid'"):
assert_array_equal(results_ols.df_resid, results_ols.df_residuals)
with pytest.warns(FutureWarning, match="'resid'"):
assert_array_equal(results_ols.resid, results_ols.residuals)
with pytest.warns(FutureWarning, match="'wY'"):
assert_array_equal(results_ols.wY, results_ols.whitened_Y)
with pytest.warns(FutureWarning, match="'wdesign'"):
assert_array_equal(results_ols.wdesign, results_ols.whitened_design)
with pytest.warns(FutureWarning, match="'wdesign'"):
assert_array_equal(model.wdesign, model.whitened_design)
def test_OLS_degenerate():
Xd = X.copy()
Xd[:, 0] = Xd[:, 1] + Xd[:, 2]
model = OLSModel(design=Xd)
results = model.fit(Y)
assert results.df_residuals == 31
def test_OLS():
model = OLSModel(design=X)
results = model.fit(Y)
assert results.df_residuals == 30
assert results.residuals.shape[0] == 40
assert results.predicted.shape[0] == 40
"""
import numpy as np
# In fact we're testing methods defined in model
from nistats.regression import OLSModel
from nose.tools import assert_true, assert_equal, assert_raises
from nose import SkipTest
from numpy.testing import (assert_array_almost_equal, assert_array_equal)
N = 10
X = np.c_[np.linspace(-1, 1, N), np.ones((N, ))]
Y = np.r_[range(5), range(1, 6)]
MODEL = OLSModel(X)
RESULTS = MODEL.fit(Y)
""" R script
::
X = cbind(0:9 * 2/9 -1, 1)
Y = as.matrix(c(0:4, 1:5))
results = lm(Y ~ X-1)
print(results)
print(summary(results))
gives::
Call:
lm(formula = Y ~ X - 1)
def test_OLS():
model = OLSModel(design=X)
results = model.fit(Y)
assert_equal(results.df_resid, 30)