How to control for confounds in decoding analyses of neuroimaging data ("MultiVoxel Confound Analysis", MVCA)
Code for simulations and empirical analyses for our article on dealing with confounds in decoding analyses of neuroimaging data. The data (preprocessed VBM/TBSS arrays) can be downloaded using the download_data.py script.
We've implemented the (cross-validated) confound regression procedure as a scikit-learn-style "transformer" class, which is implemented in the skbold package. To use it, install skbold (from master) and import it as follows: from skbold.preproc import ConfoundRegressor.
NEW: we also uploaded the source code for the ConfoundRegressor object in this repository at analyses/confounds.py,
so no need to install skbold (feel free to use/modify/adapt/etc).
A minimal example on how to use it is outlined below:
import numpy as np
from skbold.preproc import ConfoundRegressor
data = np.random.normal(0, 1, size=(100, 2))
# X = neuroimaging data (N x K array), C = confound (N x P array, for P confound variables)
X, C = data[:, 0, np.newaxis], data[:, 1]
X_train, X_test = X[::2, :], X[1::2, :]
cfr = ConfoundRegressor(confound=C, X=X)
X_train_corr = cfr.fit_transform(X_train)
X_test_corr = cfr.transform(X_test)
# X_train_corr and X_test_corr represent the data corrected for the confoundThe ConfoundRegressor class also works with scikit-learn pipelines!
import numpy as np
from skbold.preproc import ConfoundRegressor
from sklearn.pipeline import make_pipeline
from sklearn.svm import SVC
from sklearn.model_selection import cross_val_score
data = np.random.normal(0, 1, size=(100, 2))
# X = neuroimaging data (N x K array), C = confound (N x P array, for P confound variables)
X, C = data[:, 0, np.newaxis], data[:, 1]
y = np.repeat([0, 1], repeats=50)
cfr = ConfoundRegressor(confound=C, X=X)
clf = SVC(kernel='linear')
pipeline = make_pipeline(cfr, clf)
scores = cross_val_score(estimator=pipeline, X=X, y=y, cv=10)