Empirical Likelihood in Nonignorable Covariate-Missing Data Problems.

Area of Science:

• Statistics
• Biostatistics
• Econometrics

Background:

• Missing covariate data is a common challenge in regression analysis across health, social sciences, and survey sampling.
• Nonignorable missingness requires specialized methods beyond complete case analysis to avoid biased results.
• Existing semiparametric approaches utilize working probability and score models for nonignorable missing covariates.

Purpose of the Study:

• To develop an empirical likelihood approach for analyzing nonignorable covariate-missing data problems.
• To effectively integrate two working models (missingness probability and conditional score) for robust estimation.
• To propose a unified system of unbiased estimating equations for efficient parameter estimation.

Main Methods:

• Developed a unified approach to construct unbiased estimating equations, incorporating incomplete data.
• Applied empirical likelihood methodology to optimally combine these estimating equations.
• Proposed three maximum empirical likelihood estimators and compared their finite-sample performance via simulation.

Main Results:

• The proposed empirical likelihood method offers efficient estimation of regression parameters.
• Simulation studies demonstrate superior performance in terms of bias, efficiency, and robustness compared to competitors.
• The method was successfully illustrated using data from the US National Health and Nutrition Examination Survey (NHANES).

Conclusions:

• Empirical likelihood provides a powerful and flexible framework for handling nonignorable missing covariate data.
• The proposed estimators are efficient and robust to potential model misspecifications.
• This method enhances the analysis of complex datasets common in health and social science research.