It's biasing the sample. For example, drug usage among teens are roughly the same between white, hispanic and african americans but african americans are more likely to be arrested and charged for drug possession or usage. Here the difference is probably because african americans are stopped and searched more even though their usage is about the same as whites.