just some thoughts
do we use joblib?
Alexandre rewrote the permutation t-test to use joblib.
Problems on Windows
see especially Robert's comments
cost of creating new process on Windows:
my guess: if each new process needs to import scipy.stats, then there is not much point to multiprocessing on Windows.
correct or not? (example would be an rvs from scipy.stats.distributions)
added joblib in #111
importing joblib from scikit-learn is slow
needs at least a few seconds of calculations to pay for the extra cost
REF: numdiff add lower bound on epsilon, closes #72