Assessing the Performance of EM Algorithm and Multiple Imputation in Beta Regression with Missing Data

Abstract

Handling missing data remains a fundamental challenge in statistical modeling, particularly within regression models. This study evaluates and contrasts two widely used imputation techniques, the Expectation-Maximization (EM) algorithm and Multiple Imputation (MI), in the context of beta regression. The EM algorithm iteratively estimates missing values by maximizing the likelihood function, while MI generates multiple plausible datasets to account for the uncertainty of missing data. Using the gasoline yield dataset with artificially induced missingness at 5%, 10%, and 15%, we assessed the performance of both methods across various link functions and likelihood estimators. Findings suggest that while both methods are effective at lower missingness levels, EM consistently yields more robust parameter estimates at moderate levels of missingness (around 10%), and maintains strong performance as it increases, especially when coupled with the log-log link function. These findings may offer valuable insights for researchers and practitioners dealing with incomplete data in beta regression models.