Multiple Imputation of Missing Data in Longitudinal Designs: A Comparison of Different Strategies

Missing data frequently occur in longitudinal designs and are commonly addressed using multiple imputation (MI), either in the form of multilevel MI, which treats repeated measures as nested within participants, or single-level MI, which treats repeated measures as separate variables. Previous research has shown that both approaches can perform well in applications of latent curve models (LCMs) in which the assumptions underlying multilevel MI are met. Using two simulation studies, we extend prior work in two directions: (1) to applications of single-indicator LCMs in which the assumptions of multilevel MI are violated, and (2) to multiple-indicator designs, in which multilevel MI can be challenging to implement and single-level MI may be computationally unstable due to the large number of variables. Results indicate that single-level MI provides the most accurate results in the contexts studied here, especially in conditions in which common implementations of multilevel MI are misspecified, and that large numbers of variables can be accommodated in single-level MI by employing dimension reduction techniques such as partial least squares. We also discuss the implications of these findings for applied research and illustrate the application of these methods with an empirical example.