Multiple Imputation of Nonnormal Level 2 Missingness in Multilevel Models

• In Event: 45.075 - Multilevel Models, Multiple Imputation, and Latent Models: Unrelated but Equally Important Topics

Abstract

Nested data structure and missing data are common situations in social science. This study aimed to extend Grund, Lüdtke, & Robitzsch (2017) by investigating how the three multilevel based multiple imputation methods (namely joint modeling and fully conditional specification with manifest vs. latent means) handle nonnormal level 2 missingness as opposed to normal data while utilizing HLM as the analysis model rather than simplistic linear regression. The results demonstrated how these imputation methods were affected by sample size, and more importantly, their sensitivity to violation of normality at level 2. Even though the performances deteriorated across the board with nonnormal level-2 missingness, joint modeling outperformed the other two methods. Implications of these results for empirical researchers are discussed as well.

Authors

• Yutang Xiong, University of Denver

• Duan Zhang, University of Denver