Multiple bias calibration for valid statistical inference under nonignorable nonresponse.

Area of Science:

• Statistics
• Biostatistics
• Survey Methodology

Background:

• Nonresponse bias poses significant challenges to valid statistical inference.
• Accurate statistical analysis requires robust methods to handle missing data and potential biases.

Purpose of the Study:

• To develop a novel statistical method for addressing nonresponse bias.
• To eliminate selection bias in statistical inference using multiple propensity score models and empirical likelihood.

Main Methods:

• The proposed method utilizes multiple candidate propensity score (PS) models within an empirical likelihood framework.
• Multiple bias calibration is introduced, incorporating multiple working PS models into the empirical likelihood's internal bias calibration constraint.
• Asymptotic properties of the method are investigated.

Main Results:

• Selection bias can be safely eliminated under specific conditions: the true model must be among the working PS models, and their expectations must match the true missing rate.
• The method demonstrates theoretical advantages in handling nonresponse bias.
• Simulation studies compare the proposed method against existing techniques.

Conclusions:

• Multiple bias calibration offers a robust approach to mitigate nonresponse bias in statistical inference.
• The method provides a practical solution for real-world data analysis, as demonstrated by its application to body fat percentage data.