Power Analysis for Latent Growth Models With Nonignorable Dropout

• In Event: Sharp Statistics in Educational Research

Abstract

Recently, longitudinal data have been increasingly used to assess students’ growth in educational research. Various missingness mechanisms need to be considered in an experimental design to determine an appropriate sample size for desirable power. The current study compares two Monte Carlo methods to study power in latent growth models with ignorable and nonignorable dropout. Based on our findings, the Yuan, Zhang, and Zhao (2017) method is preferred over the Muthen and Muthen (2002) because the Type I error rates for the former are well controlled throughout the simulation conditions. We also illustrate the simulation-based guidelines to demonstrate how the number of measurement occasions and amount of missing data can affect power to detect significant variability in individual growth trajectories.

Authors

• Seong eun Hong, University of Massachusetts - Amherst

• Scott Monroe, University of Massachusetts - Amherst