Misclassification of a prognostic dichotomous variable: sample size and parameter estimate adjustment.

Under general conditions, Lagakos showed that for an explanatory variable observed with error, the asymptotic relative efficiency (ARE) when using the observed rather than the true values in linear models, logistic models and proportional hazards models for survival is the square of the correlation between the true and observed variables. The result is useful for sample size adjustment when this correlation is estimable. Often, one cannot observe correct values of the explanatory variable under any circumstances. We show, however, that under the models considered by Lagakos for a dichotomous explanatory variable, the ARE equals the kappa statistic in a read-reread protocol. Consequently, one need not know 'truth' in this situation to estimate the ARE and to adjust sample size to maintain desired power; divide the estimated sample size obtained with the assumption of no measurement error by the consistent estimate of the kappa statistic (which is unlikely to be zero or negative). We then develop heuristically an adjusted estimate of the beta parameter in a proportional hazards survival model. The work was motivated by analyses of the Childhood Brain Tumour Consortium database. Examples from this database illustrate the method.