The Impact of Correlated Outcomes and Missing Data on Power in Multilevel Designs

• In Event: Power and Design, Model Misspecification, and Multilevel Modeling

Abstract

The purpose of this study is to systematically explore the relationship between power and between-outcomes correlations for tests of group mean differences when both complete and incomplete multivariate data are collected from individuals who are nested within clusters. Using Monte Carlo simulation, we compared the power of the multivariate multilevel model (MVMM), which simultaneously estimates group mean differences and outcome correlations, to that of the strictly univariate hierarchical linear model (HLM), which does not take these correlations into account. We varied five factors: outcome correlations, standardized effect sizes, percentage of missing outcome data, intraclass correlation, and sample size. Preliminary results indicate that use of the MVMM yielded power levels similar to or greater than those obtained from conducting separate HLMs.

Authors

• Wanchen Chang, Boise State University

• Keenan A. Pituch, The University of Texas - Austin