Use and limitations of dual measurements in correcting for nondifferential exposure misclassification

Recently, several authors have encouraged the use of dual measurement strategies to correct for bias due to nondifferential misclassification in epidemiologic research. Among these, latent class techniques, which give unbiased results if both measurements are independent conditional on the true value, have become most popular. In practice, however, measurement errors are usually more likely to be positively correlated, and hence reliability studies cannot replace studies of validity. I offer here a quantitative illustration and a comparison of the performance of latent class approaches and other dual response strategies in situations of positive error covariance. I conclude that, except under very special circumstances, dual measurement strategies are likely to remove only part, if any, of the effect attenuation due to nondifferential exposure misclassification. Under certain conditions, it may make sense, however, to estimate a "minimum" effect of the exposure on the basis of the strongest association found in various dual and single measurement strategies.