Statistical analysis for categorical data subject to misclassification errors

点击翻译按钮获取中文摘要

The proposal aims to tackle some emerging issues in analyzing categorical data that are not perfectly assessed. The primary goal is to (1) enrich the toolbox of statistical methodologies on categorical data subject to misclassification errors and (2) provide valid and insightful guidelines for applied practitioners. Categorical data subject to misclassification errors has been commonly encountered in many applied fields such as health science, psychology, and genetic studies. For instance, smoking status is measured by a self-report questionnaire. Due to the fear of stigmatization, participants may not always report their true smoking behaviour. It has been long recognized that simply ignoring such errors can lead to biased inference that may severely distort the truth. Most of the existing literature focus on either only misclassified covariates (usually binary) or only misclassified response variables (usually binary). Rather limited work has been done for dealing with simultaneous misclassification of response variables and covariates, which has been becoming increasingly common in real world. For example, studies that share the same sources of information (e.g., self-report questionnaire) to assess both the response variable (e.g. disease status) and the covariate (e.g., smoking status) tend to have correlated misclassification errors. This proposal aims to tackle the challenges arising from the real applications in three different yet interconnected directions: (1) developing models that can accommodate correlated misclassification of the response variable and the covariate; (2) assessing consequences of using incorrect misclassification models; (3) developing test procedures for diagnosing the validity of assumptions about the misclassification structure.