An approach to categorical data with nonignorable nonresponse

Area of Science:

• Statistics
• Biostatistics
• Statistical Modeling

Background:

• Log-linear models are utilized for categorical outcomes with nonignorable nonresponse.
• Existing methods like maximum likelihood can yield unstable boundary estimates.
• Park and Brown (1994) proposed a pseudo-Bayesian method to smooth unobserved cell frequencies.

Purpose of the Study:

• To propose a generalized pseudo-Bayesian approach for nonignorable nonresponse.
• To compare the proposed method with Park and Brown's approach and maximum likelihood estimation.
• To enhance the stability of estimates in log-linear models with missing categorical data.

Main Methods:

• Fitting log-linear models to augmented frequency tables.
• Implementing a pseudo-Bayesian method assigning prior observations to both observed and unobserved cells.
• Conducting a simulation study for comparative analysis.

Main Results:

• The proposed approach demonstrates improved performance over maximum likelihood.
• The new method offers a generalization of Park and Brown's technique.
• Simulation results indicate enhanced stability and accuracy in parameter estimation.

Conclusions:

• The generalized pseudo-Bayesian approach provides a robust alternative for handling nonignorable nonresponse in categorical data.
• This method addresses limitations of maximum likelihood, particularly boundary estimation issues.
• The study contributes a more stable and effective statistical tool for analyzing incomplete categorical datasets.