X Matters Too: A Case for the Ubiquity of Multilevel Models

• In Event: Applications in Multilevel Modeling

Abstract

This study systematically investigated how dependencies due to cluster membership in level 1 predictors, in the absence of cluster dependencies in the outcome, affect regression coefficient estimates and coverage when cluster membership is ignored. A 2-level, 2-predictor model was simulated with varying degrees of X ICCs, R2, X-X correlations, clusters, and cluster sizes. Results show a negative effect of X ICC on unilevel predictor slope coefficient estimates, ranging from -7% to -33% (X ICC = .10 to .40), which corresponds to a blend of the within- and between-slope values. Coverage estimates were similarly poor, ranging from 84% to 32% (X ICC = .10 to .40). Importantly, use of unilevel cluster-robust standard errors does not fix the problem. Implications are discussed.

Authors

• Elizabeth A. Sanders, University of Washington

• Timothy R. Konold, University of Virginia