The effect of misclassification by one-, two-, or seven-day food records on the apparent magnitude of results was quantified for two hypothetical models of association of diet with disease. For each of 106 women, classification to quintiles of intake was calculated from 37 to 72 one-day records and compared to classification by one- to seven-day records. In analyses based on few records per subject, odds ratios were biased toward unity and results from models differing in strength of association of diet with disease were more similar. Loss of power in a test for trend was especially important for associations of the magnitude probable for diet and disease relationships (odds ratios of 3.00 or less) and for samples of 100 cases and 100 controls or fewer. The measurement error associated with diet measures currently in use can obscure relatively strong associations even without biased or confounded measures.