It seems to me that if we reject a subset of experimental samples because they look like bad data (e.g. extreme outlier caused by sensor malfunction) we are still keeping all the bad data we are unable to recognize as such (e.g. sensor malfunctions producing less extreme data), which introduces a bias.