Crossdressers, transgender people of various stripes, etc may not be all that abundant in a strictly numerical sense, but we are damn near guarunteed to throw an exception to your criteria every time you ask one of us.

Your criteria account perfectly for the majority, up until they encounter an admitted exception and then almost invariably fail. Those exceptions aren't noise in the dataset -- they're a sign that you're ignoring the points that don't neatly fit your aesthetically-pleasing line.