Statistical Methods to Adjust for Misclassified Repeated Exposures in Modeling Disease-Exposure Associations

In public health studies, it is common for exposure status to be misclassified. In this dissertation, statistical models to adjust for misclassification will be proposed to address four related questions of interest.The first question focuses on exploring the association between a disease and the unobservable probability of true exposure given error-prone exposure replicates, in a case-control setting when the exposure is a binary variable. Assuming a beta distribution for the exposure probability, we obtain the estimated association by maximizing the marginal likelihood of the observed replicates and disease status.The second question is motivated by the same study, but our interest shifts to assessing the relationship between a disease and the true unknown binary exposure status. We generalize a regular latent class model and a latent class model with a random effect to incorporate a true disease model, for situations of both conditionally independent and dependent exposure replicates. A real data example from the Baltimore Washington Infant Study will be presented to demonstrate methods addressing the first two questions.In general ANOVA settings, when the samples are misclassified in the study leading to an incorrect attribution of group membership, appropriate adjustments are necessary to obtain valid estimates and inferences. As a methodological transition into our next motivating example, we address this general misclassification problem as our third question, with focus on adapting both the classic regression calibration method and the likelihood method to correct for misclassification in this setting utilizing external or internal validation data.Our fourth question aims to identify whether a subject's true mean and/or variability in exposure exceeding certain thresholds bears any association with a disease outcome. Misclassifications arise in the categorization of whether the continuous mean exposure or variance exceeds a relevant threshold, where the true mean or variance itself is unobserved. Methods to be discussed include derivations based on matrix method, regression calibration, a full likelihood approach, and a two-stage empirical Bayes method incorporating categorizations based on both exposure means and variances. Simulation results and analysis of exposure and outcome data from the Mount Sinai Study of Women Office Workers will be presented.