A Comparison of Some Bootstrap Methods for Mixed Model Analysis of Non-Normal Longitudinal Data

• Activity Number: 170 - Nonparametric Methods for Longitudinal and Survival Data
• Type: Contributed
• Date/Time: Monday, July 31, 2017 : 10:30 AM to 12:20 PM
• Sponsor: Section on Nonparametric Statistics
• Author(s): Mark Reiser*
• Companies: Arizona State University
• Keywords: cluster bootstrap ; non-parametric bootstrap ; missing data ; simulation ; mixed-model ; longitudinal data

Abstract

A longitudinal study may have clustering at more than one level. For example, in a longitudinal study of school children, observations on the same student over time produce clustering at the child level, and observations on children from the same classroom (teacher) produce clustering at the classroom level. Clustering at more than one level leads to complications in calculating bootstrap estimates of confidence intervals. In this study, confidence intervals from the non-parametric bootstrap with case-resampling are studied for parameters of both fixed effects and variance components under a mixed model with clustering at two levels applied to non-normal data with missing values. For a one-level bootstrap, a case is defined as all observations from the same higher-level unit. A two-level bootstrap is also studied. Using Monte Carlo simulations, the study will examine the effect of number of clusters, cluster size, and imputation model on validity of the bootstrap confidence intervals. Results are given for the bootstrap-t, percentile, and percentile-BCa methods. An application to an educational study is also presented.