Comparing Bayesian and Likelihood-Based Approaches in Multilevel Models With Few Clusters and Nonnormal Data

• In Event: AERA Poster Session 4
  In Poster Session: Division D Section 2 Poster Session 1

Abstract

This study aims to compare the performance of the Bayesian and restricted maximum likelihood (REML) estimation method in the analysis of multilevel models with small numbers of clusters and non-normal data. Simulated data were used to examine the effects of cluster size, non-normality of data, prior distributions on the parameter estimates of a two-level multilevel model. Preliminary results showed that the severe non-normality led to biased estimates of Level-2 regression coefficients for both approaches, but the Bayesian estimation method generated less biased estimates of fixed effects than the REML method. The standard error estimates of Level-2 predictor coefficients were unbiased for both Bayesian and REML. Lastly, the Bayesian approach yielded larger level-2 residual variance estimates than REML.

Author

• lanrong li, Amplify Education, Inc.