A Bayesian Latent Variable Selection Model for Nonignorable Missingness.

Area of Science:

• Statistics
• Data Science
• Behavioral Sciences

Background:

• Missing data are prevalent in educational, social, and behavioral sciences.
• Missing Not At Random (MNAR) mechanisms, where missingness relates to unobserved data, pose significant challenges.
• Existing MNAR methods often result in incomplete data and struggle with complex covariate structures (interactions, non-linear terms, random slopes).

Purpose of the Study:

• To propose a novel Bayesian latent variable imputation approach for handling missing data, including MNAR mechanisms.
• To simultaneously estimate the model of substantive interest alongside data imputation.
• To address limitations of existing methods by accommodating complex covariate structures.

Main Methods:

• Developed a Bayesian latent variable imputation approach.
• Simultaneously estimated the imputation model and the substantive model.
• The proposed Bayesian Latent Variable Selection Model (BLVSM) was evaluated through computer simulations.

Main Results:

• BLVSM demonstrated effectiveness in imputing data with MNAR mechanisms, yielding results comparable to complete data analysis in most scenarios.
• Performance was particularly strong with large sample sizes.
• Compared to methods assuming Missing At Random (MAR), BLVSM provided less biased estimates and better coverage under MNAR conditions.
• Performance was less satisfactory under specific conditions with small sample sizes and high missingness proportions.

Conclusions:

• The proposed BLVSM is a robust method for handling missing data, particularly MNAR, in complex datasets.
• It offers an improvement over traditional MAR-based methods, providing more accurate estimates.
• The method is practical for real-world applications and available in the Blimp software.