Bayesian adjustment for exposure misclassification in case-control studies

Measurement error occurs frequently in observational studies investigating the relationship between exposure variables and the clinical outcome. Error-prone observations on the explanatory variable may lead to biased estimation and loss of power in detecting the impact of an exposure variable. The mechanism of measurement error, such as whether or in what way the quality of data is affected by the disease status, is seldom completely revealed to the investigators. This increases uncertainty in assessing the consequences of ignoring measurement error associated with observed data, and brings difficulties to adjustment for mismeasurement. In this study, we consider situations with a correctly specified binary response, and a misclassified binary exposure. We propose a solution to conduct Bayesian adjustment to correct for measurement error subject to varying differentiality, including the nondifferential misclassification, differential misclassification and nearly nondifferential misclassification. Our Bayesian model incorporates the randomness of exposure prevalences and misclassification parameters as prior distributions. The posterior model is constructed upon simulations generated by Gibbs sampler and Metropolis-Hastings algorithm. Internal validation data is utilized to insure the resulting model is identifiable. Meanwhile, we compare the Bayesian model with maximum likelihood estimation (MLE) and simulation extrapolation (MC-SIMEX) methods, using simulated datasets. The Bayesian and MLE models produce accurate and similar estimates for odds ratio in describing the association between the disease and exposure, when appropriate assumptions regarding the differentially of misclassification are made. The 90% credible or confidence intervals capture the truth approximately 90% of the time. A Bayesian method corresponding to nearly nondifferential prior belief compromises between the loss of efficiency and loss of accuracy associated with other prior assumptions. At the end, we look at two case-control studies with misclassified exposure variables, and aim to make valid inference about the effect parameter.Science, Faculty ofStatistics, Department ofGraduat