Misclassification of covariates.

In the context of reported data on asthma mortality, we examine two types of covariate misclassification. The first type is non-differential misclassification, in which the proportion of subjects misclassified is invariant over exposure and disease status. Building on work by Cox and Elwood and by Blettner and Wahrendorf, we find that the range of admissible values for misclassification proportions is bounded by the observed data, and may not include any values that account for observed heterogeneity of effect estimates. The second type is differential misclassification, in which the classification error differs according to the disease or exposure classes to which study subjects belong. If the relation between exposure and the confounding variable or between the confounding variable and disease is strong, differential misclassification can produce large variations in the stratum-specific odds ratio estimates.