Bias Analysis for Logistic Regression with a Misclassified Multi-categorical Exposure

Abstract

In epidemiological studies, it is one common issue that the collected data may not be perfect due to technical and/or nancial di culties in reality. It is well known that ignoring such imperfections may lead to misleading inference results (e.g., fail to detect the actual association between two variables). Davidov et al.(2003) have studied asymptotic biases caused by misclassi cation in a binary exposure in a logistic regression context. The aim of this thesis is to extend the work of Davidov et al. to a multi-categorical scenario. I examine asymptotic biases on regression coe cients of a logistic regression model when the multicategorical exposure is subject to misclassi cation. The asymptotic results may provide insight guide for large scale studies when considering whether bias corrections would be necessary. To better understand the asymptotic results, I also conduct some numerical examples and simulation studies.