Date of Award
Doctor of Philosophy (PhD)
Michelle L. Kelley
James F. Paulson
Mediation and moderated mediation models are two commonly used models for indirect effects analysis. In practice, missing data is a pervasive problem in structural equation modeling with psychological data. Multiple imputation (MI) is one method used to estimate model parameters in the presence of missing data, while accounting for uncertainty due to the missing data. Unfortunately, commonly used MI methods are not equipped to handle categorical variables or nonlinear variables such as interactions. In this study, we introduce a general MI framework that uses the Bayesian bootstrap (BB) method to generate posterior inferences for indirect effects and gradient boosted machine learning imputation models that can impute missing data in linear and logistic regression models with linear and nonlinear effects.
Two Monte Carlo simulation studies are conducted to examine the empirical performance of a BB procedure for estimation and inference of indirect effects and to examine the performance of the proposed imputation algorithm in indirect effects analysis. Results show that the BB has comparable performance to widely used frequentist methods (e.g., delta methods and nonparametric bootstrap with bias-correction) for indirect effects analysis for a variety of models and conditions. With missing data, in general, results indicate that the proposed MI framework has comparable performance to model-based estimation and other MI algorithms for indirect effects analysis in mediation models; for indirect effects analysis in moderated mediation models, results indicate that the proposed MI framework outperforms these methods in most conditions. Advantages and limitations of the BB as applied to indirect effects analysis are discussed.
Milletich, Robert J. II, "Multiple Imputation of Missing Data in Structural Equation Models with Mediators and Moderators Using Gradient Boosted Machine Learning" (2016). Psychology Theses & Dissertations. 44.