Chapter 15: Missing Data Handling for Multilevel Data

A substantial body of methodological research supports techniques that assume a missing at random (MAR) mechanism, whereby the missing data indicator for an incomplete variable (e.g., 0 = complete, 1 = missing) is unrelated to that variable’s scores after conditioning on observed data (Little & Rubin, 2002; Rubin, 1976). Maximum likelihood estimation and multiple imputation are two such methods that enjoy widespread use, and Bayesian estimation is a third approach gaining in popularity (van de Shoot et al., 2017). Despite the literature’s preference for MAR-based techniques, multilevel software packages often force the user to remove cases with missing data, thereby assuming a missing completely at random (MCAR) mechanism where missingness is unrelated to the analysis variables (i.e., no systematic predictors of nonresponse). Deletion has been characterized as “among the worst methods available for practical applications” (Wilkinson & Taskforce on Statistical Inference, 1999, p. 598), and its effects can be particularly devastating with incomplete level-2 variables because the entire cluster is excluded from analysis, even if the corresponding level-1 variables are complete.