Assessing bias in community-based prevalence estimates: towards an unduplicated count of problem drinkers and drug users.