Small Sample Performance of Multilevel and Traditional Methods for Multivariate Group Comparisons With Incomplete Data

• In Event: 34.094 - Fairmont Royal York Roundtable Session Six
  In Roundtable Session: 34.094-5 - Partially Clustered Data and Sample Size Investigations With Various Types of Multilevel Models

Abstract

A simulation study was conducted to compare the performance of multivariate multilevel models (MVMM) and conventional multivariate methods to assess two-group mean differences with multiple outcomes under small sample and missingness conditions. When total sample size was 40 or larger, the likelihood ratio test with MVMM provided accurate Type I error rates and the greatest power for the omnibus multivariate test. Under these same conditions, MVMM conventional t tests provided accurate Type I error rates and the greatest power for tests of specific outcomes. For extremely small samples, the MVMM adjusted likelihood ratio test (α=.025) and the t test using the Kenward-Roger correction had the best performance, with the Kenward-Roger correction also providing accurate confidence interval coverage across all conditions.

Authors

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

• Tiffany Ann Whittaker, The University of Texas at Austin

• Megha Joshi, The University of Texas at Austin

• Ryoungsun Park, SiFive

• Molly Elizabeth Cain, The University of Texas at Austin