Count regression models with a misclassified binary covariate : a Bayesian approach.

Mismeasurment, and specifically misclassification, are inevitable in a variety of regression applications. Fallible measurement methods are often used when infallible methods are either expensive or not available. Ignoring mismeasurement will result in biased estimates for the associated regression parameters. The models presented in this dissertation are designed to correct this bias and yield variance estimates reflecting the uncertainty that is introduced by flawed measurements. We consider a generalized linear model for a Poisson response. This model accounts for the misclassification associated with the binary exposure covariate. In the first portion of the analysis, diffuse priors are utilized for the regression coefficients and the effective prior sample size technique is implemented to construct informative priors for the misclassification parameters. In the second portion of the analysis we place informative priors on the regression parameters and diffuse priors on the misclassification parameters. We also present results of a simulation study that incorporates prior information for both the regression coefficients and the misclassification parameters. Next, we extend the Poisson model with a single binary covariate in various ways, including adding a continuous covariate and accounting for clustering through the use of random effects models. We also consider a zero-inflated version of the model. Simulation studies are summarized for each extension. Finally, we discuss an application in which frequentist and Bayesian logistic regression models are used to predict prevalence of high BMI-for-age among preschool-aged children in Texas.