UCLA

Over the last few decades, a large body of research supports the use of multiple imputation as a method for handling missing data. Despite imputation’s broad appeal, the method is known to introduce biases when applied to models with interactive and polynomial effects. In the context of single-level regression models, multiple imputation based on a fully Bayesian model specification has shown great promise, but limited research to date has considered this approach for multilevel models. The purpose of this dissertation is to investigate the multilevel extension of Bayesian model-based imputation to a two-level regression model with a cross-level interactive effect.With the exception of some rather extreme scenarios with non- normal data, computer simulations from this research suggest that the model-based approach can effectively estimate these models in a wide variety of conditions that are typical of social and behavioral science research data. In virtually every condition examined, model-based imputation outperformed existing alternatives to handling incomplete interactive effects. This procedure is available in the Blimp software package for macOS, Windows, and Linux.