Which, by definition, wouldn't make it bad data.
But instead of improving the fit, it's as if the bad data were clarifying some new pattern.