Misclassification of exposure in case-control studies: assessment by quality indices.

Abstract

Misclassification of a binary exposure in case-control studies is usually assessed by sensitivity and specificity of the measured exposure. Sensitivity and specificity do not adequately measure the degree of misclassification because they assess only the proportion of truly exposed (or unexposed) subjects who are misclassified by the defective measurement. They do not account for the proportion of subjects, categorized by the defective measurement as either exposed or unexposed, who are misclassified, in other words, the predictive value of the measurement. A more appropriate way of measuring misclassification is by "quality indices" that take both of these criteria into account and that are essentially rescaled sensitivity and specificity, or predictive value, measures. I present the relation between measured and actual odds ratios in terms of quality indices. If quality indices are nondifferential or proportional, there is no misclassification bias to the odds ratio. The relation offers a new approach to correcting measured odds ratios.