Variable Selection in Multilevel Models

• In Event: AERA Poster Session Saturday 3:45 pm
  In Invited Poster Session: Division D Graduate Student In-Progress Research Poster Session

Abstract

Variable selection is essential in social science research, where data often have multilevel structures. Few studies have examined how these methods perform under multilevel data with varying conditions. This simulation study evaluates the performance of Elastic Net (Enet), glmmLasso, and Mnet across a range of ICC levels (.10–.50) and sample sizes. Data were generated from two-level multilevel models using Monte Carlo simulation, with predictors drawn from multivariate normal distributions and discretized into ordinal scales. Results provide empirical insight into how data conditions affect the performance of convex and non-convex penalized models in multilevel data. Findings aim to guide researchers in choosing appropriate penalization techniques for clustered data, improving model accuracy and interpretability in large-scale educational and social science research.

Authors

• Siyao Du, University of Texas at Austin

• Man Chen, University of Texas at Austin