When MICE and FIML Disagree: Investigating Inconsistencies in Missing Data Analysis

Missing data are pervasive and pose a significant challenge in social, behavioral, and health research. Although Multivariate Imputation by Chained Equations (MICE) is widely used because of its flexibility and practical usefulness, recent empirical studies have shown that different MICE algorithms and full information maximum likelihood (FIML) estimation may produce inconsistent statistical conclusions. This study systematically compared the performance of FIML and commonly used MICE algorithms for handling missing data in linear regression models.

Simulation studies were conducted to compare FIML with three commonly used MICE algorithms, including predictive mean matching (PMM), Bayesian linear regression (norm), and classification and regression trees (CART). Simulation conditions varied sample size, effect size, number of auxiliary variables, missing data mechanism, and missing data rate.

Results showed that although MICE is a powerful approach for handling missing data, some default settings of MICE algorithms produced biased model estimations under certain conditions. Performance also varied across the imputation algorithms and could depart from the FIML benchmark.

Findings highlight the importance of carefully selecting and implementing MICE algorithms when handling missing data and provide practical guidance for choosing between MICE and FIML in applied research.