Seeing funny behaviour where for nearly-Gaussian maps MELODIC changes the
value of `mmthresh`, making the combined maps inconsistent.
  However, the overall aim is simply to make the null of the DR maps `≈N(0,1)`.
Instead of using the full mixture model, we simply rescale the maps based on
robust estimators of the mean and standard deviation (i.e. that should be
relatively insensitive to the heavy tails).