Mirror, Mirror, on the Wall—Which Criterion Performs Best of All?

• In Event: Longitudinal Data, Model Evaluation, and Multilevel Models

Abstract

In addition to considering a reasonable theoretical rationale, researchers often use model fit statistics to compare competing statistical models. Comparative fit statistics such as information criteria are commonly used for this purpose. Multiple criteria include sample size weights on the per-parameter penalties incorporated into their computations, in order to reduce the impact of large sample sizes on model selection. However, in hierarchical linear (multilevel) modeling, a different sample size exists at each level of the statistical equation and calculation these fit statistics becomes more complex. Within this framework, researchers must decide which sample size is appropriate for computing these statistics, but prior research has presented mixed findings. The present study adds to this literature for methodologists, researchers, and practitioners.

Authors

• Sarah D. Newton, University of Connecticut

• D. Betsy Mccoach, University of Connecticut