Identifiability Conditions for Nonignorably Missing Binary Outcome Data

• Activity Number: 234
• Type: Topic Contributed
• Date/Time: Monday, August 10, 2015 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Jiwei Zhao*
• Companies: SUNY Buffalo
• Keywords: nonignorable missing data ; identifiability ; missing data mechanism ; logistic regression ; probit regression

Abstract

In epidemiology, regression models with binary outcomes are often used to investigate the relation between disease status and other exposures or covariates of interest, especially for case control studies. Clinically, these studies usually encounter the problem of missing disease status and the missing data mechanism is highly suspected to be nonignorable (Little and Rubin, 2002). Therefore, we have to be careful resolving the identifiability conditions of the unknown parameters for each method we use (Robins, 1997). In this paper, we systemically study the identifiability conditions with missing response data when the mechanism is nonignorable. Although we focus on logistic regression and probit regression, the theory can be extended to other generalized linear models. Comprehensive simulation studies and a real data analysis are conducted to illustrate our theory and method.