Novel Methods for Addressing Bias from Misclassified Exposure Variables
Abstract
Exposure variables are often misclassified in observational studies. Any analysis that does not make proper adjustments for misclassification may result in biased estimates of model parameters and that may lead to distorted inferences.
In this dissertation I study this bias in cohort and retrospective matched case-control study. For the matched case-control study, I consider a binary exposure variable whereas for the cohort study I consider a multicategory exposure variable that has more than two nominal categories. The novel aspect of this work is the use of instrumental variables to reduce the bias due to the misclassification of the exposure variable when no validation data are available. Each of the works, one involving matched case-control and the other involving the cohort data, consists of two major steps.
First I study the parameter identifiability and obtain sufficient conditions for identifiability. Then I propose model estimation and inference methods after adopting the sufficient conditions of identifiability. In the first work, I use two methods of estimation including the efficient approach. In the second work, I use a variational Bayesian inference procedure aided by the automatic differentiation variational inference (ADVI) technique. Operating characteristics of the methods are assessed and compared with existing approaches through simulation studies. Simulation studies clearly indicate when and how the proposed methods are advantageous.
Each of the methods are applied to analyze real datasets. For the matched case-control study scenario, the proposed methods are applied to the nested case-control data sampled from the 1989 United States birth registry where the reported smoking status of mothers during pregnancy is considered to be the misclassified exposure. For the cohort data scenario, the proposed Bayesian method is applied to the US breast cancer mortality data sampled from the Surveillance Epidemiology and End Results (SEER) database, where reported treatment therapy is considered to be the misclassified exposure variable.
Citation
Manuel, Christopher Matthew (2020). Novel Methods for Addressing Bias from Misclassified Exposure Variables. Doctoral dissertation, Texas A&M University. Available electronically from https : / /hdl .handle .net /1969 .1 /192562.