Measurement error as an explanation for the alcohol harm paradox: analysis of eight cohort studies.