Robust Missing Data Methods for Categorical Regression

Henley, Steven

Abstract

Improved methods for obtaining robust statistical inferences from categorical regression models in the presence of missing data and model misspecification would be an invaluable tool to the epidemiological and health care research communities. Presently epidemiological models are typically designed to identify patterns of alcohol-related symptoms, define criteria of alcohol use disorders, and evaluate policies regulating use and distribution of alcoholic beverages. Such models frequently rely on datasets that contain incomplete-data. While commercially available statistical software provides some automated missing value procedures (e.g., data imputation, Expectation-Maximization), further theoretical and empirical research is required to develop more robust statistical methods. Martingale Research will develop Robust Expectation-Maximization (REM) algorithms that combine recent advances in stochastic estimation, asymptotic statistics, and maximum likelihood recoding to achieve the following objectives: 1) design and implement REM algorithms that are suited to categorical regression modeling for epidemiological problems, 2) theoretically and empirically investigate REM algorithm performance in the presence of missing data and model misspecification and 3) extend the REM algorithms to optimally estimate preprocessing transformations in the presence of missing data and model misspecification. These results will demonstrate the essential technical feasibility required for further Phase II investigations and provide the foundation for developing commercially available software.