The Effect of Unbalanced Design and Estimation Methods in Multilevel Modeling

• In Event: Multilevel Modeling Methods and Applications II

Abstract

This simulation study examines how unbalanced designs affect parameter estimates and their standard errors in multilevel modeling. The following conditions were examined: intra-class correlation, number of groups, and unbalanced design. Two estimation procedures were used to test the accuracy of multilevel model estimation: iterated generalized least squares (IGLS) and Markov chain Monte Carlo (MCMC). Relative bias of parameter estimates and standard error were found to be sensitive to the number of groups and unbalanced design under IGLS. As the number of groups increased, the relative biases tended to decrease. With severely unbalanced design, the 95% confidence interval had substantial under-coverage. In general, use of MCMC estimation resulted in less bias across all conditions as compared to IGLS estimation.

Authors

• Hyewon Chung, Chungnam National University

• Jiseon Kim, University of Washington - Seattle

• Ryoungsun Park, The University of Texas - Austin