Reference correction introduces shifts in PSDs

[wronk]

For some HCP subjects (~1/2 of the 20 I'm testing), hcp.preprocessing.apply_ref_correction seems to be introducing significant changes in the across-sensor average PSD. This came up when trying to classifying resting state from motor trials. Resting state runs seemed to have wide-band increases in power after ref correction. Plots examples below are an average example of this effect -- some are worse, some are better.

• bad channels and reference channels were removed prior to PSD calculation for all plots below
• Annotations added to raw files using hcp.read_annots. (PSD now respects these annotations 4345)

PSD on raw unprocessed data:

PSD on raw data + BP filtering (0.1-55Hz, fir filter, 2s window):

PSD on raw data + SSP projectors (exact same projectors used for every run):

PSD on raw data + hcp.preprocessing.apply_ref_correction: