Selecting Correct Multilevel Models: Investigating Akaike and Bayesian Information Criteria and “Sample Size” Issues

• In Event: Multilevel Modeling: Estimation, Model Selection, and Sample Size

Abstract

The identification of one preferred multilevel model is a necessary, but difficult decision; researchers typically rely on theory and model fit statistics such as the Bayesian Information Criterion (BIC) to choose. In the context of multilevel models specifying random slopes, the present simulation study examines performance of the BIC when calculated with number of clusters (BIC-L2), as opposed to total number of level-1 units (BIC-L1), as compared to the Akaike Information Criteria (AIC) and chi-square test of deviances. Researchers found that for conditions featuring small numbers of clusters and total level-1 units, the AIC performs best, (favoring the true model most often) followed by the BIC-L2, the chi-square test, and finally the BIC-L1.

Authors

• Sarah D. Newton, University of Connecticut

• D. Betsy Mccoach, University of Connecticut