Author : Linying Ji
Publisher :
ISBN 13 :
Total Pages : pages
Book Rating : 4.:/5 (119 download)
Book Synopsis Missing Data in Multilevel Vector Autoregressive Model by : Linying Ji
Download or read book Missing Data in Multilevel Vector Autoregressive Model written by Linying Ji and published by . This book was released on 2020 with total page pages. Available in PDF, EPUB and Kindle. Book excerpt: Missing data are inevitable in longitudinal studies. Some missingness occurs by design, but most others are unplanned and may have profound consequences on inferential results. Numerous previous studies have evaluated different missing data handling techniques in cross-sectional and longitudinal panel data. However the impact of different kinds of missingness and strategies to handle missingness in multivariate, multilevel time-series data is less studied. It has been shown that ignoring the clustered structure in handling missing data when performing multiple imputation (MI) for cross-sectional data will lead to different variance and covariance properties of the imputed data, and thus the model estimation results are very likely to be biased (Grund, Lüdtke, & Robitzsch, 2018; van Buuren, 2018). In this dissertation project, I conducted a series of simulation studies to evaluate the performances of different multilevel MI methods in the context of Bayesian multilevel vector autoregressive models (MVARs) under different sample size, intraclass correlation, and missing data conditions. The missing data handling methods considered included a Bayesian equivalent of full-information maximum likelihood approach (BFIML), single-level MI, multilevel MI with joint modeling approach or full conditional specification approach, and a hybrid approach that integrated multilevel MI and BFIML method. Simulation results suggested that in fitting MVAR model under the parameter settings considered, researchers can effectively treat both MAR and, to a certain extent, MNAR missingness, with appropriate multilevel MI approach alone, or in combination with BFIML method. Specifically, for random-intercept only MVAR models with MAR missing data, the multilevel MI approaches performed as well as BFIML. For MNAR missingness, the multilevel MI and BFIML hybrid approach performed better. I also identified groups of parameters (i.e., random-intercept parameters) in the MVAR model that were more prone to biased estimates with MNAR missingness using the approaches considered. Finally, I discussed possible ways of resolving these limitations.